DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 2023/03/31. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1–20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding independent claim 1
Step 1 — whether the claim falls within any statutory category. See MPEP 2106.03.
Claim 1 is drawn to a method claim, and dependent claims 2–20 are likewise drawn to a method. Therefore, each of these claims falls under one of the four categories of statutory subject matter (process/method, machine/product/apparatus, manufacture, or composition of matter).
Step 2A Prong 1 — whether the claim recites a judicial exception. See MPEP 2106.04, subsection II.
Claim 1 is directed to a method for analyzing events in a system, the method comprising:
The limitations of "determining, with the processor, a plurality of event sequences from the event data, the plurality of event sequences having variable lengths," "determining, with the processor, a plurality fixed-length embeddings of the plurality of event sequences," and "performing, with the processor, a cluster analysis of the fixed-length embeddings to determine a plurality of clusters of event sequences in the plurality of event sequences" recite an abstract idea.
These limitations are directed towards the abstract idea of a mathematical relationship, specifically organizing information and manipulating information through mathematical correlations, because computing a vector representation of a sequence and grouping such vectors by similarity are mathematical calculations. See MPEP 2106.04(a)(2), subsection I, A.
These limitations are additionally directed towards the abstract idea of a mental process, specifically a concept that can be performed in the human mind, including observation, evaluation, judgement or opinion, because a human being can identify groupings of events in observed data, characterize each grouping by a set of values, and sort those characterizations into groups of similar sequences. See MPEP 2106.04(a)(2), subsection III. The recitation that these operations are performed "with the processor" does not remove them from the mental-process grouping.
The limitation of "receiving, with a processor, event data from the system" is evaluated as an additional element in Step 2A Prong 2.
Step 2A Prong 2 — whether the claim as a whole integrates the recited judicial exception into a practical application of the exception or whether the claim is "directed to" the judicial exception. This evaluation is performed by (1) identifying whether there are any additional elements recited in the claim beyond the judicial exception, and (2) evaluating those additional elements individually and in combination to determine whether the claim as a whole integrates the exception into a practical application. See MPEP 2106.04(d).
Claim 1 recites additional elements of "a processor," "receiving, with a processor, event data from the system, the event data indicating events that occurred in the system and times at which the events occurred," and "[a] method for analyzing events in a system."
The recited processor is claimed at a high level of generality and merely acts as a tool on which the abstract operations are performed, amounting to no more than mere instructions to apply the exception on a generic computer. See MPEP 2106.05(f). The receiving step merely obtains the data on which the abstract analysis is performed and is insignificant extra-solution activity. See MPEP 2106.05(g). The recitation of analyzing events "in a system" amounts to no more than generally linking the use of the judicial exception to a particular technological environment or field of use, because the claim does not require applying the determined clusters to control, adjust, halt, or otherwise modify the operation of the monitored system. See MPEP 2106.05(h).
Further, claim 1 does not recite an improvement to the functioning of a computer or to any other technology or technical field. See MPEP 2106.04(d)(1) and 2106.05(a). Although the specification asserts advantages over prior approaches (Specification, ¶¶ [0004]–[0006]), any such advantage resides in the analytical technique itself rather than in the operation of any computer or technical field, and claim 1 does not recite the particular embedding pipelines the specification credits with producing that advantage. The claim therefore does not reflect the disclosed improvement. Considered individually and in combination, these additional elements do not integrate the exception into a practical application.
Step 2B — whether the claim amounts to significantly more than the judicial exception. See MPEP 2106.05.
The additional elements beyond the judicial exception are the recited "processor" and the receiving step. As explained above, the processor amounts to mere instructions to apply the exception on a generic computer (MPEP 2106.05(f)) and the receiving step is insignificant extra-solution activity (MPEP 2106.05(g)). Upon reevaluation in Step 2B, court decisions discussed in MPEP 2106.05(d)(II) recognize generic computer functions and generic data gathering as well-understood, routine, and conventional activity. See MPEP 2106.05(d) and 2106.07(a)(III)(B). Considered individually and in combination, these additional elements do not provide an inventive concept.
Accordingly, claim 1 does not recite additional elements, individually or in combination, that amount to significantly more than the recited judicial exception, and is ineligible under 35 U.S.C. § 101.
Regarding dependent claims 2–17, 19, and 20
Claims 2–17, 19, and 20 merely narrow the previously cited abstract idea limitations. For the reasons described above with respect to independent claim 1, these judicial exceptions are not meaningfully integrated into a practical application, or significantly more than the abstract ideas.
Step 2A Prong 1
Regarding claims 2, 3, and 4, these claims recite the limitations of "determining, with the processor, a chronological time series of events from the event data," "combining the multiple sets of event data into the chronological time series of events," and "labeling, with the processor, each event from the event data as a respective event type from a predetermined set of event types." These limitations are directed towards the abstract idea of a mental process. They recite arranging observations in order, merging observations from several sources, and categorizing observations under a predetermined scheme.
Regarding claim 5, this claim recites the limitations of "a first event type indicating that a measurable parameter of the system has a value that is outside of a predetermined or expected range" and "a second event type indicating that a process performed by the system is halted." These limitations are directed towards the abstract idea of a mental process. They recite evaluating whether an observed value falls outside an expected range and observing that an operation has stopped.
Regarding claims 6, 7, and 8, these claims recite forming each respective event sequence "as a subset of sequential events from the event data," such that it "begins with at least one sequential event of a first event type and ends with at least one sequential event of a second event type," and such that "a time between a last event and a first event in the respective event sequence is less than a predetermined maximum amount of time." These limitations are directed towards the abstract idea of a mental process. They recite selecting a contiguous run of observations according to stated start, stop, and duration criteria.
Regarding claims 9, 10, and 11, these claims recite "determining, with the processor, a plurality of parameter sequences," "determining a plurality of fixed-length event subsequences from the respective event sequence," "determining a plurality of fixed-length parameter subsequences," and that those subsequences "each have a same length." These limitations are directed towards the abstract idea of a mental process and a mathematical relationship. They recite compiling ordered lists of values and enumerating substrings of a stated length, which a human being can perform with pen and paper.
Regarding claims 12, 13, and 14, these claims recite "determining a first frequency vector" and "a second frequency vector" whose values "each indicate a number of occurrences of a respective possible parameter subsequence" or "event subsequence," "determining the respective fixed-length embedding of the respective event sequence as a concatenation of the first frequency vector and the second frequency vector," and vector lengths "equal to a total number possible parameter subsequences" and "a total number possible event subsequences." These limitations are directed towards the abstract idea of a mathematical relationship, specifically organizing information and manipulating information through mathematical correlations. See MPEP 2106.04(a)(2), subsection I, A.
Regarding claims 15, 16, and 17, these claims recite "determining a kernel matrix having values indicating a similarity between each possible combination of two event sequences," "determining the respective value in the kernel matrix . . . as a dot product of the respective second frequency vector and the respective third frequency vector," counting occurrences "with a predetermined number of mismatches or less," and "applying kernel principal component analysis to the kernel matrix." These limitations are directed towards the abstract idea of a mathematical relationship. A dot product and a principal component analysis are mathematical calculations. See MPEP 2106.04(a)(2), subsection I.
Regarding claims 19 and 20, these claims recite "extracting event patterns from the plurality of event sequences based on the plurality of clusters of event sequences" and "predicting, with the processor, a possible future event based on a partial event sequence and the plurality of clusters of event sequences." These limitations are directed towards the abstract idea of a mental process. They recite identifying recurring arrangements among grouped observations and forming an opinion as to what will occur next.
Step 2A Prong 2
Claims 2–17, 19, and 20 recite the additional element of "the processor." This limitation amounts to no more than mere instructions to apply the exception on a generic computer and to generally linking the use of the judicial exception to a particular technological environment or field of use. See MPEP 2106.05(f) and 2106.05(h). None of these claims recites applying the result of the analysis to control, adjust, halt, or otherwise modify the operation of the monitored system, and none recites an improvement to computer functionality or to any other technology. See MPEP 2106.04(d)(1), 2106.05(a), and 2106.05(e).
Step 2B
Considered individually and in combination, the additional elements amount to no more than using a generic processor to perform the narrowed abstract operations, which is recognized as well-understood, routine, and conventional activity. See MPEP 2106.05(d) and 2106.07(a)(III)(B). These claims do not provide an inventive concept and are ineligible under 35 U.S.C. § 101.
Regarding dependent claim 18
Step 2A Prong 1
To the extent claim 18 incorporates the limitations of claim 1, those limitations recite the abstract idea for the reasons discussed above. The limitation of "displaying, on a display screen, at least some of the plurality of clusters of event sequences" is evaluated as an additional element in Step 2A Prong 2.
Step 2A Prong 2
Claim 18 recites additional elements of "a display screen" and "displaying, on a display screen, at least some of the plurality of clusters of event sequences." The display screen is a generic computer component recited at a high level of generality, amounting to mere instructions to apply the exception on a generic computer. See MPEP 2106.05(f). The displaying step merely outputs the result of the abstract analysis and is insignificant post-solution activity. See MPEP 2106.05(g). Claim 18 does not require that anything be done in response to the displayed clusters, and therefore does not recite an improvement to computer functionality or to any other technology. See MPEP 2106.04(d)(1) and 2106.05(a).
Step 2B
Court decisions discussed in MPEP 2106.05(d)(II) recognize the presentation of information on a display and generic computer output functions as well-understood, routine, and conventional activity. See MPEP 2106.05(d) and 2106.07(a)(III)(B). Considered individually and in combination, the additional elements of claim 18 do not provide an inventive concept, and claim 18 is ineligible under 35 U.S.C. § 101.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1–4, 6–9, 18, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, and further in view of Xu et al. (Panpan Xu), US 11,074,276 B2.
Regarding Claim 1, (Jianwu Xu) teaches a method comprising:
(Jianwu Xu) teaches "A method for analyzing events in a system, the method comprising:" Specifically, (Jianwu Xu) discloses that "the present invention relates to analyzing event sequence" (Jianwu Xu, col. 1, lines 14-17).
(Jianwu Xu) teaches "receiving, with a processor, event data from the system, the event data indicating events that occurred in the system and times at which the events occurred." Specifically, (Jianwu Xu) discloses that discrete event records produced by the monitored system are received, and that "each such record can include time stamps and descriptions of the system events" (Jianwu Xu, col. 9, lines 49-51). (Jianwu Xu) further discloses, with reference to an exemplary set of received event records, that "as can be seen in this exemplary set, each record includes a time stamp and an event description" (Jianwu Xu, col. 10, lines 12-15), the exemplary records being of the form "2016/03/18 18:01:46 unix: [ID 608654 kern.notice]." These disclosures teach receiving event data from the system in which the event data indicates both the events that occurred and the times at which those events occurred.
(Jianwu Xu) teaches "determining, with the processor, a plurality of event sequences from the event data." Specifically, (Jianwu Xu) discloses that each received event record is "mapped to an identifier which denotes its event type," such that the record sequence is converted into "identifiers of event types for ease of computation" (Jianwu Xu, col. 10, lines 44-48). (Jianwu Xu) further discloses that an event corpus is generated "by taking the original sequence and performing a linear walk 502," in which the processor performs "routine 511—Start walking/traversing from each event identifier and end the walk when the same identifier is reached," "routine 512—If the same identifier type appears immediately, then walk until a different identifier type is obtained," and "routine 513—If the same identifier type is not found within a predefined number of walk, then stop the walk and restart routine 511 from the next identifier; then end the linear walk once the procedure covers all event identifiers in the sequence" (Jianwu Xu, col. 11, lines 2-15). (Jianwu Xu) further discloses that "each subsequent element in the corpus 503 represents a different event identifier sentence" (Jianwu Xu, col. 11, lines 26-29). These disclosures teach determining, from the received event data, a plurality of event sequences.
(Jianwu Xu) teaches "determining, with the processor, a plurality fixed-length embeddings." Specifically, (Jianwu Xu) discloses an encoder "configured to encode each of the event types identified in the event sequence corpus of block 372 into a d-dimensional vector representation projected onto a d-dimensional metric space" (Jianwu Xu, col. 7, lines 27-30). (Jianwu Xu) further discloses that "each event type can be represented by a d-dimensional vector with real values" (Jianwu Xu, col. 11, lines 37-39), and that the dimension d is a fixed parameter of the encoder, since "any value between 100 and 300 can be a suitable number of dimensions for the latent representation of event record types" (Jianwu Xu, col. 11, lines 1-4). Because every representation produced by the encoder has the same dimension d, these disclosures teach determining a plurality of fixed-length embeddings.
(Jianwu Xu) teaches "performing, with the processor, a cluster analysis of the fixed-length embeddings." Specifically, (Jianwu Xu) discloses a clusterer "configured to use the vector representations of a plurality of event types from block 374 to group the vector representations into clusters and to retain only those clusters which include representations of failure" event types (Jianwu Xu, col. 7, lines 30-35). (Jianwu Xu) further discloses that "process 405 involves applying a density-based clustering algorithm to automatically group the d-dimensional latent representations of event types from the output of encoder 374," and that "the DBSCAN algorithm is selected to be used in the filtering process 752 because the algorithm does not need the number of clusters to be specified a priori and can find arbitrary shaped clusters" (Jianwu Xu, col. 12, line 64 – col. 13, line 6). (Jianwu Xu) further discloses "an event type sequence clusterer 276 [that] can be configured to cluster event types and retain only clusters with failure event types" (Jianwu Xu, col. 6, lines 37-39). These disclosures teach performing a cluster analysis of the fixed-length vector representations.
Specifically, (Jianwu Xu) discloses that the linear walk terminates upon differing conditions, namely when the same identifier is reached or after a predefined number of walk steps (Jianwu Xu, col. 11, lines 6-13), and discloses encoding into a d-dimensional vector representation (Jianwu Xu, col. 7, lines 27-30). However, (Jianwu Xu) does not teach "the plurality of event sequences having variable lengths" and does not teach "determining, with the processor, a plurality fixed-length embeddings of the plurality of event sequences," because the d-dimensional vector representations of (Jianwu Xu) are representations of individual event types rather than of the event sequences, as (Jianwu Xu) states that "each event type can be represented by a d-dimensional vector" (Jianwu Xu, col. 11, lines 37-39).
In the same field of endeavor, (Kraus) teaches "the plurality of event sequences having variable lengths" and "determining, with the processor, a plurality fixed-length embeddings of the plurality of event sequences."
(Kraus) teaches "the plurality of event sequences having variable lengths." Specifically, (Kraus) discloses that the vectorizing operation is performed by "performing on the single piece of text an algorithm 1226 that learns fixed-length feature representations from variable-length pieces of text" (Kraus, col. 25, lines 4-9). (Kraus) further defines reference numeral 1226 as an "algorithm which learns fixed length feature representations (i.e., vectors) from variable length pieces of text," and reference numeral 1228 as "learn fixed length feature representations (i.e., vectors) from variable length pieces of text; may also be referred to as embedding 1106 or be part of embedding 1106" (Kraus, col. 16, lines 7-11). (Kraus) further discloses that "Lengths 808 may also be worth noting, at least to the extent of distinguishing between variable length items and fixed length items" (Kraus, col. 20, lines 56-58), and that the model "was trained 1008 using one or more ordered event sequences whose length 808 as text 804 differs from a length 808 of the extracted ordered event sequence as text," such that "the candidate sequence forms text of some different length" (Kraus, col. 22, lines 57-63). These disclosures teach a plurality of event sequences that have variable lengths.
(Kraus) teaches "determining, with the processor, a plurality fixed-length embeddings of the plurality of event sequences." Specifically, (Kraus) discloses "vectorizing 800 the candidate event sequence at least in part by embedding 1106 the candidate event sequence in a vector space, thereby producing a candidate vector" (Kraus, col. 24, lines 32-35). (Kraus) further discloses that "the anomaly detection approach embeds 1106 the output sequences into a vector space 810, to represent each sequence 410 as a multi-dimensional vector 1002" (Kraus, col. 28, lines 46-49), and that "[t]his embodiment constructs an account's model by feeding 1224 its event sequence documents into the doc2vec algorithm 1226. Doc2vec embeds 1106 the documents into a lower dimensional vector space 810," such that "[t]he final model 402 contains or consists of sequence vectors 814" (Kraus, col. 31, lines 7-13). (Kraus) further discloses that "Event sequences 410 extracted from logs 302 or other event lists 216 are vectorized 800 and embedded 1106 in a vector space 810" (Kraus, col. 35, lines 12-14). These disclosures teach determining, for each of a plurality of event sequences, a fixed-length embedding of that event sequence.
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and reducing the resulting event sequences to a vector representation, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-13), including "industrial process control devices 130" (Kraus, col. 19, lines 53-55). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the variable-length-sequence to fixed-length-embedding technique of (Kraus), such that the event sequences produced by the linear walk of (Jianwu Xu) are themselves embedded into fixed-length vectors.
The motivation to combine (Jianwu Xu) and (Kraus) is to preserve the sequential ordering information that (Jianwu Xu) expressly identifies as lost by its own event-type-level representation, stating that "[b]ecause averaging dilutes the event pattern ordering . . . information regarding the sequential or temporal order of the event types is often not well represented" (Jianwu Xu, col. 12, lines 24-27), since (Kraus) teaches that embedding whole event sequences yields a "similarity function [that] captures sequence similarity semantics: sequences are similar if they share similar events that appear in a similar order" (Kraus, col. 28, lines 53-55).
The combination of (Jianwu Xu) and (Kraus), however, does not teach "performing, with the processor, a cluster analysis of the fixed-length embeddings to determine a plurality of clusters of event sequences in the plurality of event sequences." Specifically, the cluster analysis of (Jianwu Xu) groups the d-dimensional latent representations of event types (Jianwu Xu, col. 12, line 64 – col. 13, line 2), rather than determining a plurality of clusters of event sequences, and (Kraus) computes an anomaly score for a candidate vector using a k-nearest neighbors calculation (Kraus, col. 31, lines 18-27) rather than determining a plurality of clusters of event sequences.
In the same field of endeavor, (Panpan Xu) teaches "performing, with the processor, a cluster analysis of the fixed-length embeddings to determine a plurality of clusters of event sequences in the plurality of event sequences." Specifically, (Panpan Xu) discloses "generating, with the processor, a plurality of clusters using a minimum description length (MDL) optimization process, each cluster in the plurality of clusters including a set of at least two event sequences in the plurality of event sequences that maps to a pattern in each cluster, the pattern in each cluster having a plurality of events included in at least one event sequence in the set of at least two event sequences in the cluster" (Panpan Xu, col. 19, lines 25-33; see also col. 2, lines 17-24). (Panpan Xu) further discloses that the input to the clustering operation is a plurality of event sequences, namely "receiving, with a processor, a plurality of event sequences, each event sequence in the plurality of event sequences including a plurality of events" (Panpan Xu, col. 19, lines 22-25). (Panpan Xu) further discloses that the resulting clusters partition the input sequences, since the merging operation "ensures that each event sequence in the original input appears only once in the final summarized output event sequence" (Panpan Xu, col. 9, lines 51-53). These disclosures teach performing a cluster analysis that determines a plurality of clusters of event sequences within the plurality of event sequences.
(Jianwu Xu), (Kraus), and (Panpan Xu) are analogous to the claimed invention as all three are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and grouping the resulting event sequences to identify patterns associated with system faults, since (Panpan Xu) is directed to "[e]vent sequence data, i.e., multiple series of timestamped or ordered events," including "vehicle error logs in automotive industry" (Panpan Xu, col. 1, lines 20-24). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu), as modified by the sequence-level fixed-length embedding of (Kraus), with the sequence-level cluster analysis of (Panpan Xu), such that the cluster analysis operates on the fixed-length embeddings and produces a plurality of clusters each containing a set of event sequences.
The motivation to combine (Jianwu Xu), (Kraus), and (Panpan Xu) is as recited by (Panpan Xu), namely that grouping whole event sequences resolves the "visual 'clutter' due to the noisy and complex nature of the event sequences with high event cardinality" (Panpan Xu, col. 1, lines 48-53) and yields "typical fault development paths," which "can inform better strategies to prevent the faults from occurring" (Panpan Xu, col. 1, lines 33-35), which is the same fault-diagnosis objective pursued by (Jianwu Xu), namely to "help pinpoint the causes of system faults and failures" (Jianwu Xu, col. 2, lines 46-48).
As to dependent Claim 2, the claim recites:
The method according to claim 1 further comprising:
determining, with the processor, a chronological time series of events from the event data,
wherein the plurality of event sequences is determined from the time series of events.
Regarding the limitation "The method according to claim 1 further comprising:", Claim 2 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
However, (Jianwu Xu) does not teach those limitations mentioned above.
In the same field of endeavor, (Kraus) teaches "determining, with the processor, a chronological time series of events from the event data" and "wherein the plurality of event sequences is determined from the time series of events."
(Kraus) teaches "determining, with the processor, a chronological time series of events from the event data." Specifically, (Kraus) discloses that "Input to an algorithm 402 is derived from a system's event log 302, which documents a time-series of events 204 that occurred in the system 130. Each record (row) 204 in the event log corresponds to a single event and contains event information such as the operation performed, parameters, and some metadata" (Kraus, col. 28, lines 8-13). (Kraus) further discloses that "[t]he storage log contains a time series of events 204, where each event record contains event features, e.g., timestamp, operation type, and error code" (Kraus, col. 30, lines 17-20). (Kraus) further discloses that the time series so obtained is chronologically ordered, stating that "[a] log is often ordered chronologically, but a composite log may include events from machines whose timestamps are not fully synchronized, and thus may be only partially ordered" (Kraus, col. 19, lines 62-65). These disclosures teach determining, from the event data, a time series of events that is ordered chronologically.
(Kraus) teaches "wherein the plurality of event sequences is determined from the time series of events." Specifically, (Kraus) discloses that "one approach uses a heuristic extraction 502 that breaks the log into multiple sequences, where each sequence approximates 1242 a user session 506. An output is a set of sequences 410, each of which was likely performed by a single user 104" (Kraus, col. 28, lines 22-26). (Kraus) further discloses that "one extraction 502 partitions the log into event sequences, with each sequence originated by the same IP address and application, each sequence consisting of no more than T subsequent events, and where no more than alpha seconds have passed between any two consecutive events in the sequence" (Kraus, col. 28, lines 29-35). (Kraus) further discloses that "a sequence 410 represents an entity's consecutive events for a period of time," and that "[h]euristics 502 can be applied to separate sessions by using idle time and the number of events as delimiters 1222" (Kraus, col. 32, lines 33-36), giving the worked example in which "the 10 minute gap shown below as a comment would result in recognition 502 of two event sequences, before and after the gap respectively, instead of a single sequence containing all the events shown" (Kraus, col. 32, lines 36-40), the illustrated events being listed by chronological timestamp from 12:04 through 13:24 (Kraus, col. 32, lines 41-52). Because the log that (Kraus) partitions into the plurality of event sequences is the same chronologically ordered time series of events described at Kraus, col. 28, lines 8-13 and col. 30, lines 17-20, these disclosures teach that the plurality of event sequences is determined from the time series of events.
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and forming those records into event sequences, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-13). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the chronological time-series construction and log-partitioning of (Kraus), such that the event data of (Jianwu Xu) is first assembled into a chronological time series of events from which the plurality of event sequences is then determined.
The motivation to combine (Jianwu Xu) and (Kraus) is as recited by (Kraus), namely that partitioning a chronologically ordered event log yields sequences each corresponding to a single coherent unit of activity, since (Kraus) teaches that the extraction produces "a set of sequences 410, each of which was likely performed by a single user 104" (Kraus, col. 28, lines 25-26), and that "these delimiters can be enforced in a computationally very efficient manner, allowing efficient heuristic extraction 502" (Kraus, col. 28, lines 38-41).
As to dependent Claim 3, the claim recites:
The method according to claim 2, wherein the event data includes multiple sets of event data from multiple sources of event data, the method further comprising:
combining the multiple sets of event data into the chronological time series of events.
Regarding the limitation "The method according to claim 2," Claim 3 depends from Claim 2, which in turn depends from Claim 1, and Claim 3 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1 and Claim 2 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
(Jianwu Xu) teaches "wherein the event data includes multiple sets of event data from multiple sources of event data." Specifically, (Jianwu Xu) discloses "a system and method for contextual event sequence analysis of system failure that analyzes heterogeneous system event record logs to help pinpoint the causes of system faults and failures and track their spread through the system's components through time during different phases of the system's operation" (Jianwu Xu, col. 2, lines 45-50). (Jianwu Xu) further discloses that "one or more discrete event records (e.g., event sequence records), such as records produced by an ICT, a manufacturing plant system, an interconnected sensor system, and the like, can be received or obtained from a record source 322 by pattern extractor 370" (Jianwu Xu, col. 9, lines 45-49). (Jianwu Xu) further discloses that "the event corpus includes multiple event sequences" (Jianwu Xu, col. 3, lines 20-21), and that the analysis tracks how a fault propagates "through time and across different components" of the monitored system (Jianwu Xu, col. 3, lines 33-36). Because the event record logs analyzed by (Jianwu Xu) are heterogeneous and are produced by different components of the monitored system, including software logging utilities and physical sensors (Jianwu Xu, col. 1, lines 11-13), these disclosures teach that the event data includes multiple sets of event data from multiple sources of event data.
(Jianwu Xu) teaches something related to "combining the multiple sets of event data into the chronological time series of events." However, (Jianwu Xu) does not teach "combining the multiple sets of event data into the chronological time series of events." The time-stamp-based ordering operation of (Jianwu Xu) is applied to the automata constructed downstream of the clustering operation (Jianwu Xu, col. 14, lines 22-24, "sorting the time series in a temporal order"), rather than to the multiple sets of received event data themselves for the purpose of forming the chronological time series of events from which the plurality of event sequences is determined.
In the same field of endeavor, (Kraus) teaches this limitation.Specifically, (Kraus) discloses, with reference to FIG. 3, that "[i]llustrated examples [of event list sources] include logs 302, e.g., syslog format logs, event tracing logs, application logs, logs generated by kernels, transaction logs, and other records of events which occurred or report state of one or more machines 102. A log is often ordered chronologically, but a composite log may include events from machines whose timestamps are not fully synchronized, and thus may be only partially ordered" (Kraus, col. 19, lines 58-65). (Kraus) further discloses that "[i]llustrated examples of event list sources 214 also include sniffers 304, which is defined broadly above to include more than merely network analyzers, and SIEM tools 306" (Kraus, col. 19, line 66 – col. 20, line 2), and defines an event list source as "computing technology, not people; may be, e.g., an event generator, or a storage location holding generated events" (Kraus, col. 12, lines 22-24). (Kraus) further discloses that the resulting event log so assembled "documents a time-series of events 204 that occurred in the system 130," where "[e]ach record (row) 204 in the event log corresponds to a single event" (Kraus, col. 28, lines 8-13), and that this time series of events is thereafter partitioned into the plurality of event sequences (Kraus, col. 28, lines 22-35). Because the composite log of (Kraus) is expressly formed of events drawn from multiple machines and is chronologically ordered, and because that composite log constitutes the time series of events from which the event sequences are extracted, these disclosures teach combining the multiple sets of event data into the chronological time series of events.
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and forming those records into event sequences, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-13). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the heterogeneous event record analysis of (Jianwu Xu) with the composite-log construction of (Kraus), such that the multiple sets of event data received by (Jianwu Xu) are combined into a single chronological time series of events.
The motivation to combine (Jianwu Xu) and (Kraus) is to enable the analysis to span the entire monitored system rather than any single component, since (Kraus) teaches a "composite log [that] may include events from machines whose timestamps are not fully synchronized" (Kraus, col. 19, lines 62-65), and (Jianwu Xu) seeks to track the spread of faults "through the system's components through time" and "across different components" (Jianwu Xu, col. 2, lines 46-50; col. 3, lines 35-36), which requires that events originating at different components be placed in a common chronological ordering.
As to dependent Claim 4, the claim recites:
The method according to claim 1 further comprising:
labeling, with the processor, each event from the event data as a respective event type from a predetermined set of event types.
Regarding the limitation "The method according to claim 1 further comprising:", Claim 4 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
(Jianwu Xu) teaches "labeling, with the processor, each event from the event data as a respective event type." Specifically, (Jianwu Xu) discloses that "as part of the process 402, once the unique event patterns are obtained, each event record is mapped to an identifier which denotes its event type. In this manner, a sequence of event records is translated into a sequence of identifiers of event types for ease of computation and manipulation in the subsequent steps of the exemplary method" (Jianwu Xu, col. 10, lines 44-50). (Jianwu Xu) further discloses that this mapping operation is performed by a processor-implemented component, namely the "pattern extractor 370" (Jianwu Xu, col. 10, lines 50-51), and that the pattern extractor together with the corpus generator, vector encoder, event type sequence clusterer, automaton generator, and sorter "can be entirely hardware, entirely software or including both hardware and software elements" executed by a computing device (Jianwu Xu, col. 7, lines 1-7). (Jianwu Xu) further discloses that "the process of extracting event patterns 402 . . . includes obtaining unique event patterns to categorize them into categories of unique event patterns/types (hereinafter event 'pattern' and event 'type' may be used interchangeably)" (Jianwu Xu, col. 10, lines 28-33). These disclosures teach labeling, with a processor, each event from the event data as a respective event type.
(Jianwu Xu) teaches that the event type is assigned "from a predetermined set of event types." Specifically, (Jianwu Xu) discloses that the event patterns to which each record is mapped "can be extracted from the records themselves or obtained from elsewhere (e.g., a known set of record types created by a software platform)" (Jianwu Xu, col. 10, lines 33-35). (Jianwu Xu) further discloses an example of such a predetermined set, stating that "in a cloud computing system using a cloud management platform, the event records can have types of events such as 'system boot', 'starting instance', 'compute instance' as well as other unique types of event that can be obtained through the domain knowledge of the underlying system or platform" (Jianwu Xu, col. 10, lines 35-41). (Jianwu Xu) further discloses that the extraction process produces a defined set of event types, namely "a set of event patterns/types which only exist during system failure periods," and that "[t]his set of event types can be denoted as the set of seed failure event types and can serve as the input in the corpus generation 403" (Jianwu Xu, col. 10, lines 51-58). Because the known set of record types created by the software platform and the types obtained through domain knowledge of the underlying system are established in advance of the mapping operation, and because each event record is thereafter mapped to an identifier drawn from that set, these disclosures teach labeling each event from the event data as a respective event type from a predetermined set of event types.
As to dependent Claim 6, the claim recites:
The method according to claim 1, the determining the plurality of event sequences further comprising:
forming each respective event sequence in the plurality of event sequences as a subset of sequential events from the event data.
Regarding the limitation "The method according to claim 1, the determining the plurality of event sequences further comprising:", Claim 6 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 6, (Jianwu Xu) teaches "forming each respective event sequence in the plurality of event sequences as a subset of sequential events from the event data."
(Jianwu Xu) teaches that each respective event sequence in the plurality of event sequences is formed "from the event data." Specifically, (Jianwu Xu) discloses that "the process of generating 403 an event corpus can be performed via a corpus generator 372. This process includes taking an event identifier sequence 501 as an input and producing event corpus 503 composed of multiple sequences (analogous to sentences) of identifiers wherein a sentence represents a sequence of identifiers (analogous to words)" (Jianwu Xu, col. 10, lines 59-66). The input event identifier sequence 501 is itself derived from the received event data, since (Jianwu Xu) discloses that "each event record is mapped to an identifier which denotes its event type. In this manner, a sequence of event records is translated into a sequence of identifiers of event types" (Jianwu Xu, col. 10, lines 44-48).
(Jianwu Xu) teaches "forming each respective event sequence . . . as a subset of sequential events." Specifically, (Jianwu Xu) discloses that "each identifier sequence (sentence) is generated by starting from any event pattern in the original sequence and performing a linear walk 502" (Jianwu Xu, col. 10, line 66 – col. 11, line 2), and sets forth the detailed procedure by which the walk selects a contiguous run of events from the original sequence: "For each event identifier in the sequence, perform routine 511—Start walking/traversing from each event identifier and end the walk when the same identifier is reached; perform routine 512—If the same identifier type appears immediately, then walk until a different identifier type is obtained; and perform routine 513—If the same identifier type is not found within a predefined number of walk, then stop the walk and restart routine 511 from the next identifier; then end the linear walk once the procedure covers all event identifiers in the sequence" (Jianwu Xu, col. 11, lines 5-15). (Jianwu Xu) further confirms that the events so selected are sequential and contiguous within the original sequence, disclosing that "two event identifiers can belong to the same cluster if one can walk from the first event identifier to the second by a predetermined 'sufficiently small' step (e.g., a step one event identifier long, a step two event identifiers long) which can be defined as the minimum traversal distance between event identifiers along an event identifier sequence" (Jianwu Xu, col. 11, lines 15-22). (Jianwu Xu) further discloses that the operation is repeated so as to produce a plurality of such sequences, since "each subsequent element in the corpus 503 represents a different event identifier sentence" (Jianwu Xu, col. 11, lines 26-29). Because the linear walk begins at a selected event identifier within the original sequence, proceeds along that sequence one event identifier at a time, and terminates upon a defined stopping condition, the resulting event identifier sentence is composed of fewer than all of the events in the original sequence and comprises only events that are sequential to one another within that original sequence.
As to dependent Claim 7, the claim recites:
The method according to claim 6, the determining the plurality of event sequences further comprising:
forming each respective event sequence in the plurality of event sequences such that the respective event sequence begins with at least one sequential event of a first event type and ends with at least one sequential event of a second event type.
Regarding the limitation "The method according to claim 6, the determining the plurality of event sequences further comprising:", Claim 7 depends from Claim 6, which in turn depends from Claim 1, and Claim 7 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1 and Claim 6 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 7, (Jianwu Xu) teaches "forming each respective event sequence in the plurality of event sequences such that the respective event sequence begins with at least one sequential event of a first event type and ends with at least one sequential event of a second event type."
(Jianwu Xu) teaches that the respective event sequence "begins with at least one sequential event of a first event type." Specifically, (Jianwu Xu) discloses that "each identifier sequence (sentence) is generated by starting from any event pattern in the original sequence and performing a linear walk 502" (Jianwu Xu, col. 10, line 66 – col. 11, line 2), and that the walk procedure begins by "Start walking/traversing from each event identifier" (Jianwu Xu, col. 11, lines 6-7). (Jianwu Xu) further discloses that each such event identifier denotes an event type, since "each event record is mapped to an identifier which denotes its event type" (Jianwu Xu, col. 10, lines 44-46), and that "event 'pattern' and event 'type' may be used interchangeably" (Jianwu Xu, col. 10, lines 31-33). The event identifier from which the walk starts therefore constitutes at least one sequential event of a first event type with which the respective event sequence begins.
(Jianwu Xu) teaches that the respective event sequence "ends with at least one sequential event of a second event type." Specifically, (Jianwu Xu) discloses that "perform routine 512—If the same identifier type appears immediately, then walk until a different identifier type is obtained" (Jianwu Xu, col. 11, lines 8-10). Routine 512 thus expressly terminates the walk upon obtaining an identifier type that is different from the identifier type at which the walk began, such that the resulting event identifier sentence ends with at least one sequential event of a second event type that is distinct from the first event type. (Jianwu Xu) further discloses that the walk otherwise terminates upon a defined type-based stopping condition, namely "end the walk when the same identifier is reached" (Jianwu Xu, col. 11, lines 6-8), confirming that the terminus of each event identifier sentence is determined by the event type of the event at which the walk stops rather than by any type-independent criterion.
As to dependent Claim 8, the claim recites:
The method according to claim 6, the determining the plurality of event sequences further comprising:
forming each respective event sequence in the plurality of event sequences such that a time between a last event and a first event in the respective event sequence is less than a predetermined maximum amount of time.
Regarding the limitation "The method according to claim 6, the determining the plurality of event sequences further comprising:", Claim 8 depends from Claim 6, which in turn depends from Claim 1, and Claim 8 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1 and Claim 6 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 8, (Jianwu Xu) teaches something related to "forming each respective event sequence in the plurality of event sequences such that a time between a last event and a first event in the respective event sequence is less than a predetermined maximum amount of time."
However, (Jianwu Xu) does not teach "forming each respective event sequence in the plurality of event sequences such that a time between a last event and a first event in the respective event sequence is less than a predetermined maximum amount of time." In the same field of endeavor, (Kraus) teaches this limitation.
(Kraus) teaches this limitation literally. Specifically, (Kraus) discloses that "[h]euristically extracting 502 the candidate event sequence may include delimiting 1222 the candidate event sequence, based on at least one of the following sequence delimiting parameters: a limit 514 on the maximum number of events allowed in the candidate event sequence, or a limit 516 on the maximum time allowed between any two consecutive events in the candidate event sequence, or a limit 516 on the maximum time allowed between an earliest event in the candidate event sequence and a latest event in the candidate event sequence" (Kraus, col. 24, lines 61-67 – col. 25, line 3). (Kraus) further discloses the same limitation in the context of the selection operation, stating that "the difference between timestamps of the earliest and latest selected events is no more than max-time-between-events 516" (Kraus, col. 22, lines 6-8), and defines the anchor event as the "event from which a hyperparameter 518 (e.g., number of events 514 or maximum time span from first event to last event) is calculated" (Kraus, col. 13, lines 55-58).
(Kraus) further teaches that the maximum amount of time is predetermined, disclosing that "heuristically extracting 502 the candidate event sequence includes enforcing at least one of the following sequence delimiting 1222 parameters 1020 . . . a limit 516 on the maximum time allowed between an earliest event in the candidate event sequence and a latest event in the candidate event sequence is in the range from three seconds to three minutes" (Kraus, col. 26, lines 61-67 – col. 27, line 6), and that "the sequence anomalies detection code 408 includes a max-time-between-events limit 516 which is in a range from one nanosecond to ten minutes" (Kraus, col. 21, lines 54-56). (Kraus) further discloses a worked example of the operation, stating that "[h]euristics 502 can be applied to separate sessions by using idle time and the number of events as delimiters 1222. For example, in some embodiments the 10 minute gap shown below as a comment would result in recognition 502 of two event sequences, before and after the gap respectively, instead of a single sequence containing all the events shown" (Kraus, col. 32, lines 34-40), the illustrated events running from "12:04, Acquire Lease" through "13:24, List Blobs" (Kraus, col. 32, lines 41-52).
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and forming those records into event sequences, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-13). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the time-span delimiting parameter of (Kraus), such that each event sequence formed by (Jianwu Xu) is delimited so that the time between the last event and the first event of that sequence is less than a predetermined maximum amount of time.
The motivation to combine (Jianwu Xu) and (Kraus) is as recited by (Kraus), namely that delimiting each sequence by elapsed time yields sequences each representing "an entity's consecutive events for a period of time" (Kraus, col. 32, lines 33-34) while remaining inexpensive to extract, since "these delimiters can be enforced in a computationally very efficient manner" (Kraus, col. 28, lines 38-41), and since (Jianwu Xu) already bounds its sequences by a predetermined limit to capture only patterns appearing "within a short traversal distance of each other" (Jianwu Xu, col. 11, lines 22-26).
As to dependent Claim 9, the claim recites:
The method according to claim 1, wherein the event data further includes a respective parameter associated with each event in the event data, the method further comprising:
determining, with the processor, a plurality of parameter sequences, each parameter sequence being formed from the respective parameters associated with events in a respective event sequence from the plurality of event sequences,
wherein the plurality of fixed-length embeddings is determined based on the plurality of event sequences and plurality of parameter sequences.
Regarding the limitation "The method according to claim 1 . . . the method further comprising:", Claim 9 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 9, (Jianwu Xu) teaches something related to "wherein the event data further includes a respective parameter associated with each event in the event data." However, (Jianwu Xu) does not teach:
"wherein the event data further includes a respective parameter associated with each event in the event data,"
"determining, with the processor, a plurality of parameter sequences, each parameter sequence being formed from the respective parameters associated with events in a respective event sequence from the plurality of event sequences,"
"wherein the plurality of fixed-length embeddings is determined based on the plurality of event sequences and plurality of parameter sequences."
In the same field of endeavor, (Kraus) teaches "wherein the event data further includes a respective parameter associated with each event in the event data." Specifically, (Kraus) discloses that "[e]ach record (row) 204 in the event log corresponds to a single event and contains event information such as the operation performed, parameters, and some metadata. For example, a storage event log of interest contains an operation type (e.g., read data), operation parameters (e.g., data size), an error or result code, and metadata such as IP address and storage account identifier" (Kraus, col. 28, lines 10-16). (Kraus) further discloses that "[t]he storage log contains a time series of events 204, where each event record contains event features, e.g., timestamp, operation type, and error code" (Kraus, col. 30, lines 18-20), and that each event is characterized by "an event's feature tuple 806 <timestamp, authentication type, operation type, error code, account name, IP address, user agent, and response size>" (Kraus, col. 30, lines 45-48).
(Kraus) teaches "determining, with the processor, a plurality of parameter sequences, each parameter sequence being formed from the respective parameters associated with events in a respective event sequence from the plurality of event sequences." Specifically, (Kraus) discloses that "[t]his embodiment represents an event sequence 410 by a textual document 804: it transforms an event's feature tuple 806 . . . into a sequence of corresponding tokens 812, in which each token represents a feature value" (Kraus, col. 30, lines 44-49). (Kraus) further discloses a worked example of the resulting document, stating that "consider a sequence document which includes: 'read_data /dir1/dir2/file3 100-Bytes success list_files /dir1/dir5 failure 0-files list_files /dir1/dir2 success 10-files'. This represents a sequence composed of three consecutive events of read_data, list_files, and list_files and their respective parameters and metadata" (Kraus, col. 28, lines 63-66 – col. 29, line 2). The recited document accordingly contains, alongside the run of three operation-type tokens forming the event sequence, a corresponding run of parameter tokens, namely "100-Bytes," "0-files," and "10-files," each associated with a respective one of the three events of that event sequence. These disclosures teach determining a plurality of parameter sequences, each formed from the respective parameters associated with the events in a respective event sequence.
(Kraus) teaches "wherein the plurality of fixed-length embeddings is determined based on the plurality of event sequences and plurality of parameter sequences." Specifically, (Kraus) discloses that "[t]his approach feeds the sequence documents into the doc2vec algorithm, which models them as vectors and learns the desired similarity function" (Kraus, col. 28, lines 59-62), and that the algorithm so applied is one "that learns fixed-length feature representations from variable-length pieces of text" (Kraus, col. 25, lines 6-9). Because the sequence documents so fed to the algorithm contain both the tokens representing the events of the event sequence and the tokens representing the respective parameters associated with those events (Kraus, col. 28, lines 63-66 – col. 29, line 2; col. 30, lines 44-49), the fixed-length vector produced for each sequence is determined based on both the event sequence and the parameter sequence. (Kraus) confirms this result, disclosing that "[t]he final model 402 contains or consists of sequence vectors 814" (Kraus, col. 31, lines 11-13).
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system, forming those records into event sequences, and reducing the event sequences to a vector representation for further analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-10), including "industrial process control devices 130" (Kraus, col. 19, lines 53-55). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the parameter-inclusive sequence document construction of (Kraus), such that the event data received by (Jianwu Xu) further includes a respective parameter associated with each event, such that a plurality of parameter sequences is determined therefrom, and such that the fixed-length embeddings are determined based on both the event sequences and the parameter sequences.
The motivation to combine (Jianwu Xu) and (Kraus) is to retain, in the vector representation, the per-event parameter information that (Jianwu Xu) discards when it reduces each event record to a bare event type identifier "for ease of computation" (Jianwu Xu, col. 10, lines 46-49), since (Kraus) teaches that including the parameter tokens in the sequence document yields an embedding in which "the learnt similarity considers the context in which events occur, which can be beneficial in assessing sequence similarity" (Kraus, col. 29, line 64 – col. 30, line 1).
As to dependent Claim 18, the claim recites:
The method according to claim 1 further comprising:
displaying, on a display screen, at least some of the plurality of clusters of event sequences.
Claim 18 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 18, (Jianwu Xu) teaches something related to "displaying, on a display screen, at least some of the plurality of clusters of event sequences." However, (Jianwu Xu) does not teach "displaying, on a display screen, at least some of the plurality of clusters of event sequences."
In the same field of endeavor, (Panpan Xu) teaches this limitation.
(Panpan Xu) teaches displaying at least some of the plurality of clusters of event sequences. Specifically, (Panpan Xu) discloses "generating, with the processor and a display output device, a graphical depiction of a first cluster in the plurality of clusters, the graphical depiction including (i) a graphical depiction of a first plurality of events in the pattern of the first cluster and (ii) a graphical indicator of the correction for the pattern" (Panpan Xu, col. 19, lines 37-42). Because each such cluster "includ[es] a set of at least two event sequences in the plurality of event sequences" (Panpan Xu, col. 19, lines 27-29), the graphical depiction so generated is a depiction of a cluster of event sequences. (Panpan Xu) further discloses depicting more than one such cluster, stating that the method further comprises "generating, with the processor and the display output device, the graphical depiction of at least two clusters in the plurality of clusters, the graphical depiction including a graphical depiction for each of the plurality of events included in each pattern of the at least two clusters" (Panpan Xu, col. 19, lines 43-49), thereby teaching the display of at least some of the plurality of clusters.
(Panpan Xu) teaches that the depiction is displayed on a display screen. Specifically, (Panpan Xu) discloses that the system includes "a display output device 154" as a component (Panpan Xu, col. 4, line 42), and that the processor is configured to operate that device to render the depiction, whether locally or remotely (Panpan Xu, col. 6, line 55 – col. 7, line 15). (Panpan Xu) further frames the entire disclosed method as directed to the generation of such a display, since the invention is a "method for generating a graphical depiction of summarized event sequences" (Panpan Xu, col. 19, lines 20-21).
(Jianwu Xu) and (Panpan Xu) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and grouping the resulting event sequences to identify patterns associated with system faults, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Panpan Xu) is directed to "[e]vent sequence data, i.e., multiple series of timestamped or ordered events," including "vehicle error logs in automotive industry" (Panpan Xu, col. 1, lines 20-24). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence analysis of (Jianwu Xu) with the display operation of (Panpan Xu), such that at least some of the plurality of clusters of event sequences is displayed on a display screen.
The motivation to combine (Jianwu Xu) and (Panpan Xu) is as recited by (Panpan Xu), namely that displaying the clusters rather than the raw event sequences resolves the visual clutter that arises when large sets of event sequences are presented, since (Panpan Xu) teaches that the direct display of such data "often produces a visual 'clutter' due to the noisy and complex nature of the event sequences with high event cardinality, which presents challenges to constructing concise yet comprehensive overviews for such data" (Panpan Xu, col. 1, lines 48-53), and that the resulting display permits identification of "typical fault development paths," which "can inform better strategies to prevent the faults from occurring" (Panpan Xu, col. 1, lines 33-35). One of ordinary skill in the art would have been motivated to make this combination because (Jianwu Xu) already provides a display device (Jianwu Xu, col. 5, lines 40-46) and pursues the same fault-diagnosis objective, namely to "help pinpoint the causes of system faults and failures" (Jianwu Xu, col. 2, lines 46-48), which objective requires that the analytical result be presented to a human analyst.
As to dependent Claim 19, the claim recites:
The method according to claim 1 further comprising:
extracting event patterns from the plurality of event sequences based on the plurality of clusters of event sequences.
Claim 19 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 19, (Jianwu Xu) teaches something related to "extracting event patterns from the plurality of event sequences based on the plurality of clusters of event sequences." However, (Jianwu Xu) does not teach "extracting event patterns from the plurality of event sequences based on the plurality of clusters of event sequences."
In the same field of endeavor, (Panpan Xu) teaches this limitation.
(Panpan Xu) teaches that a respective event pattern is extracted for each cluster of event sequences. Specifically, (Panpan Xu) discloses "generating, with the processor, a plurality of clusters using a minimum description length (MDL) optimization process, each cluster in the plurality of clusters including a set of at least two event sequences in the plurality of event sequences that maps to a pattern in each cluster, the pattern in each cluster having a plurality of events included in at least one event sequence in the set of at least two event sequences in the cluster" (Panpan Xu, col. 19, lines 25-32). The pattern so generated is thus composed of events drawn from the event sequences belonging to that cluster, and is therefore extracted from the plurality of event sequences based on the plurality of clusters.
(Panpan Xu) teaches that the extraction of the patterns and the formation of the clusters are performed together as a single operation. Specifically, (Panpan Xu) discloses that "[t]he method enables simultaneous sequence clustering and pattern extraction and it is highly tolerant to noises such as missing or additional events in the data" (Panpan Xu, col. 2, lines 6-8). (Panpan Xu) further discloses the formal relationship between the clusters and the patterns, stating that "[t]he mapping f clusters the event sequences together: sequences that map to the same pattern P can be considered to be in a single cluster. The cluster is denoted as a tuple c = (P, G) where G = {S | S ∈ S ∧ f(S) = P} is the set of sequences mapped to the pattern P" (Panpan Xu, col. 4, lines 11-15), and that "each pattern summarizes one or more of the input sequences S" (Panpan Xu, col. 4, lines 27-29).
(Jianwu Xu) and (Panpan Xu) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and grouping the resulting event sequences to identify patterns associated with system faults, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Panpan Xu) is directed to "[e]vent sequence data, i.e., multiple series of timestamped or ordered events," including "vehicle error logs in automotive industry" (Panpan Xu, col. 1, lines 20-24). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence analysis of (Jianwu Xu) with the cluster-based pattern extraction of (Panpan Xu), such that event patterns are extracted from the plurality of event sequences based on the plurality of clusters of event sequences.
The motivation to combine (Jianwu Xu) and (Panpan Xu) is as recited by (Panpan Xu), namely that extracting a pattern from each cluster yields a summary of the underlying event sequences that reduces visual clutter while retaining the information content of the data, since (Panpan Xu) teaches that its method "addresses a fundamental trade-off in visualization design: reducing visual clutter vs. increasing the information content" (Panpan Xu, col. 2, lines 3-5), and that "each pattern summarizes one or more of the input sequences S to reduce visual clutter" (Panpan Xu, col. 4, lines 28-29). One of ordinary skill in the art would have been motivated to make this combination because (Jianwu Xu) likewise seeks to identify the event patterns that characterize a system failure, stating that its objective is to "describe the features of a system failure in different phases of a system's operation" (Jianwu Xu, col. 10, lines 23-25), and because extracting the patterns from the clusters of whole event sequences, rather than from the raw records in advance of clustering, yields patterns that characterize entire fault development paths rather than isolated event types.
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, and further in view of Mishra et al. (Mishra), US 11,017,321 B1.
As to dependent Claim 5, the claim recites:
The method according to claim 1, wherein the events of the event data include events of:
a first event type indicating that a measurable parameter of the system has a value that is outside of a predetermined or expected range; and
a second event type indicating that a process performed by the system is halted.
(Jianwu Xu) teaches "wherein the events of the event data include events of . . . a second event type indicating that a process performed by the system is halted" by disclosing an exemplary set of received event records that includes the record "2016/03/18 18:01:46 panic cpu16/thread=ffffff0003ea3c60: k_fatal 0x0212C000," which record identifies a specific executing thread together with a fatal kernel panic condition and which (Jianwu Xu) expressly classifies as a distinct type of record, stating that "[t]he second record corresponds to a failure, but it is embedded among event records corresponding to normal operation of the system" (Jianwu Xu, col. 10, lines 3-16).
However, (Jianwu Xu) does not teach "a first event type indicating that a measurable parameter of the system has a value that is outside of a predetermined or expected range."
In the same field of endeavor, (Mishra) teaches this limitation. Specifically, (Mishra) discloses that "the operating characteristics data 136 may indicate operating characteristics or sensor readings associated with the equipment asset 150, such as temperatures, pressures, vibrations, and the like," and that "the events 110 may be detected based on the operating characteristics data 136" (Mishra, col. 9, lines 49-55). (Mishra) further discloses that "the events 110 may include a temperature of a particular component exceeding a temperature threshold or a differential pressure between two components failing to satisfy a pressure threshold, as non-limiting examples" (Mishra, col. 9, lines 57-61). (Mishra) further discloses the detection operation by which such events are identified, stating that "the monitoring device 102 may compare the extracted features . . . to one or more thresholds to detect the events 110. As a non-limiting example, the monitoring device 102 may detect (or receive indication of detection of) a drill temperature event based on a determination that a temperature of a drill included in the equipment asset 150 exceeds a temperature threshold" (Mishra, col. 18, lines 38-46), and that "the expert system 204 may compare particular parameter values extracted from the operating characteristics data to various thresholds and, based on the comparisons, detect occurrence of one or more events. As a particular, non-limiting example, the expert system 204 may detect a valve overheating event based on a value of a valve temperature included in the operating characteristics data exceeding a temperature threshold" (Mishra, col. 23, lines 38-46). (Mishra) further discloses that such threshold-defined events constitute distinct categories of events, since "[t]he categorization engine 124 may group the events 110 . . . into clusters of different categories" (Mishra, col. 10, lines 3-5). Because a measured temperature exceeding a temperature threshold, and a measured differential pressure failing to satisfy a pressure threshold, each constitute a measurable parameter of the system having a value outside of a predetermined range, these disclosures teach a first event type indicating that a measurable parameter of the system has a value that is outside of a predetermined or expected range.
(Jianwu Xu) and (Mishra) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system in order to identify events and event patterns associated with system faults. (Jianwu Xu) is directed to "Information and Communication Technology systems and manufacturing plant systems with computer software logging utilities or physical sensors" (Jianwu Xu, col. 1, lines 11-13). (Mishra) is directed to "receiving operating characteristics data associated with industrial machinery and event data indicating events detected based on the operating characteristics data" (Mishra, col. 46, lines 56-59), where the equipment asset "includes or corresponds to industrial machinery, such as an oil rig, a well, a drill, a blaster, a conveyer, a ventilator fan, a mixer, a crane, a generator, a compressor, a lift, a pump, a refrigerator, a packager, a production line, a furnace, a distiller, or the like," described "in a refining, mining, or manufacturing context" (Mishra, col. 15, lines 7-18). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the threshold-based event type of (Mishra), such that the event data received and labeled by (Jianwu Xu) includes events of a first event type indicating that a measurable parameter of the system has a value outside of a predetermined or expected range.
The motivation to combine (Jianwu Xu) and (Mishra) is as recited by (Mishra), namely to detect equipment conditions that precede a fault so that the fault can be prevented, since (Mishra) teaches that threshold-defined events are evaluated for "the likelihood that the events are indicative of a status of the equipment asset 150 that is a precursor to a fault" (Mishra, col. 10, lines 8-11), thereby reducing "unplanned downtime" (Mishra, col. 1, lines 37-43), such that one would supplement the event data of (Jianwu Xu) with sensor-derived parameter-threshold events, since (Jianwu Xu) expressly contemplates receiving records from "an interconnected sensor system" (Jianwu Xu, col. 9, lines 45-48) and pursues the same objective of "pinpoint[ing] the causes of system faults and failures" (Jianwu Xu, col. 2, lines 46-48).
Claims 10–12 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, and further in view of Ali et al. (Ali), Non-Patent Literature, "A k-mer Based Approach for SARS-CoV-2 Variant Identification," International Symposium on Bioinformatics Research and Applications (ISBRA), 2021, pp. 153–164, cited in the IDS filed 03/31/2023.
As to dependent Claim 10, the claim recites the following, in which the limitations shown in bold are those not taught by the combination of (Jianwu Xu) and (Kraus), and the limitations shown in plain text are those taught by that combination:
The method according to claim 9, wherein the determining the plurality of fixed-length embeddings includes determining a respective fixed-length embedding of each respective event sequence in the plurality of event sequences by:
determining a plurality of fixed-length event subsequences from the respective event sequence;
determining a plurality of fixed-length parameter subsequences from a respective parameter sequence from the plurality of parameter sequences that is associated with the respective event sequence; and
determining the respective fixed-length embedding of the respective event sequence based on the plurality of fixed-length event subsequences and the fixed-length parameter subsequences.
Claim 10 depends from Claim 9, which in turn depends from Claim 1, and Claim 10 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1 and Claim 9 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 10, (Jianwu Xu) teaches something related to "determining a plurality of fixed-length event subsequences from the respective event sequence." Specifically, (Jianwu Xu) discloses forming shorter runs of events from a longer event sequence by a bounded traversal, namely that "each identifier sequence (sentence) is generated by starting from any event pattern in the original sequence and performing a linear walk 502" (Jianwu Xu, col. 10, line 66 – col. 11, line 2), and that such a traversal proceeds by "a step one event identifier long, a step two event identifiers long" (Jianwu Xu, col. 11, lines 15-20). However, (Jianwu Xu) does not teach "wherein the determining the plurality of fixed-length embeddings includes determining a respective fixed-length embedding of each respective event sequence in the plurality of event sequences," and does not teach "a respective parameter sequence from the plurality of parameter sequences that is associated with the respective event sequence," because (Jianwu Xu) encodes individual event types rather than event sequences, stating that "each event type can be represented by a d-dimensional vector" (Jianwu Xu, col. 11, lines 37-39), and reduces each event record to a bare event type identifier before any sequence is formed, stating that "each event record is mapped to an identifier which denotes its event type" (Jianwu Xu, col. 10, lines 44-46), such that no per-event parameter is carried into the sequences.
In the same field of endeavor, (Kraus) teaches "wherein the determining the plurality of fixed-length embeddings includes determining a respective fixed-length embedding of each respective event sequence in the plurality of event sequences." Specifically, (Kraus) discloses "vectorizing 800 the candidate event sequence at least in part by embedding 1106 the candidate event sequence in a vector space, thereby producing a candidate vector" (Kraus, col. 24, lines 32-35), that the algorithm applied is one "that learns fixed-length feature representations from variable-length pieces of text" (Kraus, col. 25, lines 6-9), and that the operation is performed for each sequence of the plurality, since "the anomaly detection approach embeds 1106 the output sequences into a vector space 810, to represent each sequence 410 as a multi-dimensional vector 1002" (Kraus, col. 28, lines 46-49) and "[t]he final model 402 contains or consists of sequence vectors 814" (Kraus, col. 31, lines 11-13).
(Kraus) further teaches "a respective parameter sequence from the plurality of parameter sequences that is associated with the respective event sequence." Specifically, (Kraus) discloses that each event record "contains event information such as the operation performed, parameters, and some metadata," including "operation parameters (e.g., data size)" (Kraus, col. 28, lines 10-16), and that the embodiment "transforms an event's feature tuple 806 . . . into a sequence of corresponding tokens 812, in which each token represents a feature value" (Kraus, col. 30, lines 44-49). (Kraus) further discloses that the resulting parameter tokens are associated with the events of the respective event sequence, giving the worked example of a sequence document representing "three consecutive events of read_data, list_files, and list_files and their respective parameters and metadata" (Kraus, col. 28, line 66 – col. 29, line 2).
(Kraus) further teaches "determining the respective fixed-length embedding of the respective event sequence." Specifically, (Kraus) discloses that "[t]his approach feeds the sequence documents into the doc2vec algorithm, which models them as vectors" (Kraus, col. 28, lines 59-62), the doc2vec algorithm being the algorithm that "learns fixed-length feature representations from variable-length pieces of text" (Kraus, col. 25, lines 6-12).
(Jianwu Xu) and (Kraus) are analogous to the claimed invention as both are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system and reducing the resulting event sequences to a vector representation, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-13). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the contextual event sequence analysis of (Jianwu Xu) with the parameter-inclusive sequence document construction and fixed-length embedding of (Kraus).
The motivation to combine (Jianwu Xu) and (Kraus) is to retain the per-event parameter information that (Jianwu Xu) discards when it reduces each record to a bare event type identifier "for ease of computation" (Jianwu Xu, col. 10, lines 46-49), since (Kraus) teaches that including the parameter tokens yields an embedding in which "the learnt similarity considers the context in which events occur, which can be beneficial in assessing sequence similarity" (Kraus, col. 29, line 64 – col. 30, line 1).
The combination of (Jianwu Xu) and (Kraus), however, does not teach "determining a plurality of fixed-length event subsequences from the respective event sequence," does not teach "determining a plurality of fixed-length parameter subsequences from" the respective parameter sequence, and does not teach determining the respective fixed-length embedding "based on the plurality of fixed-length event subsequences and the fixed-length parameter subsequences." Specifically, the linear walk of (Jianwu Xu) terminates upon variable conditions and therefore produces sequences of differing lengths rather than subsequences of a fixed length, since the walk ends "when the same identifier is reached," or "until a different identifier type is obtained," or after "a predefined number of walk" (Jianwu Xu, col. 11, lines 6-13). Further, (Kraus) applies its embedding algorithm to the sequence document as a whole (Kraus, col. 28, lines 59-62) and does not disclose first decomposing that sequence into subsequences of a fixed length or deriving the embedding from such subsequences.
In the same field of endeavor, (Ali) teaches "determining a plurality of fixed-length event subsequences from the respective event sequence" and "determining a plurality of fixed-length parameter subsequences from" the respective parameter sequence. Specifically, (Ali) discloses, under the heading "3.1 k-mers Computation," that "[f]or mapping protein sequences to fixed-length vectors, it is important to preserve order information of the amino acids within a sequence. To achieve this, we use substrings (called mers) of length k. For each spike sequence, the total number of k-mers are 'N − k + 1', where N is the total number of amino acids in the spike sequence . . . and k is a user-defined parameter for the size of each mer" (Ali, p. 156, § 3.1, ¶ 1). (Ali) further discloses that "given an alphabet Σ (a finite set of symbols), we know that a spike sequence X ∈ Σ . . . we can extract sub-strings (mers) from X of length k, which we called k-mers" (Ali, p. 157, § 3.1, ¶ 2). Because the number of subsequences so extracted is N − k + 1 and each has the same length k, these disclosures teach determining, from a respective sequence, a plurality of subsequences each of a fixed length.
(Ali) further teaches determining the respective fixed-length embedding "based on the plurality of fixed-length event subsequences and the fixed-length parameter subsequences." Specifically, (Ali) discloses that "[g]iven X, k, and Σ, we have to design a frequency feature vector Φₖ(X) of length |Σ|ᵏ, which will contain the exact number of occurrences of each possible k-mer in X" (Ali, p. 157, § 3.1, ¶ 2). Because the vector so designed has a length |Σ|ᵏ that is fixed by the alphabet Σ and the parameter k, and is populated from the counts of the extracted lk-mers, this disclosure teaches determining a fixed-length embedding of the respective sequence based on the plurality of fixed-length subsequences extracted from that sequence.
(Jianwu Xu), (Kraus), and (Ali) are analogous to the claimed invention as all three are from the same field of endeavor of transforming sequences of discrete data elements into fixed-length vector representations for machine learning analysis, since (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence and parameter sequence construction of (Jianwu Xu) and (Kraus) with the k-mer subsequence extraction of (Ali).
The motivation to combine (Jianwu Xu), (Kraus), and (Ali) is as recited by (Ali), namely that extracting fixed-length subsequences preserves the ordering of symbols within a sequence, since (Ali) teaches that "it is important to preserve order information . . . To achieve this, we use substrings (called mers) of length k" (Ali, p. 156, § 3.1, ¶ 1), which addresses the deficiency (Jianwu Xu) identifies in its own representation, namely that "information regarding the sequential or temporal order of the event types is often not well represented" (Jianwu Xu, col. 12, lines 24-27).
As to dependent Claim 11, the claim recites:
The method according to claim 10, wherein the plurality of fixed-length event subsequences and the fixed-length parameter subsequences each have a same length.
Regarding Claim 11, the combination of (Jianwu Xu) and (Kraus) does not teach "wherein the plurality of fixed-length event subsequences and the fixed-length parameter subsequences each have a same length," for the reasons set forth above in the rejection of Claim 10, namely that neither reference discloses decomposing an event sequence or a parameter sequence into subsequences of a fixed length.
In the same field of endeavor, (Ali) teaches this limitation.
(Ali) teaches that the subsequences extracted from a sequence each have the same length. Specifically, (Ali) discloses that "we use substrings (called mers) of length k. For each spike sequence, the total number of k-mers are 'N − k + 1', where N is the total number of amino acids in the spike sequence . . . and k is a user-defined parameter for the size of each mer" (Ali, p. 156, § 3.1, ¶ 1). Because a single user-defined value of k governs the size of every mer extracted, every subsequence so extracted has the same length k. (Ali) further confirms that a single value of k is applied throughout, disclosing that "k is a user-defined parameter — in our experiments we use k = 9" (Ali, p. 157, § 3.1, ¶ 3), and illustrating the operation with "[a]n example of k-mers (where k = 4) is given in Figure 2" (Ali, p. 156, § 3.1, ¶ 1), which figure depicts the extraction of successive four-symbol substrings from the single sequence "MFVFVFVLPLV" (Ali, p. 157, Fig. 2).
(Ali) further teaches that the value of k so selected governs the length of the resulting representation, disclosing that the frequency feature vector Φₖ(X) has a "length |Σ|ᵏ" (Ali, p. 157, § 3.1, ¶ 2), such that a single value of k applied to two sequences over a common alphabet yields two representations of equal length.
(Jianwu Xu), (Kraus), and (Ali) are analogous to the claimed invention for the reasons set forth above in the rejection of Claim 10, namely that (Jianwu Xu) and (Kraus) are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system, and (Ali) is reasonably pertinent to the particular problem with which the inventors were concerned, namely the transformation of variable-length sequences of discrete symbols into fixed-length vector representations. See MPEP § 2141.01(a). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence and parameter sequence construction of (Jianwu Xu) and (Kraus) with the k-mer subsequence extraction of (Ali), such that a single user-defined value of k governs the extraction of the fixed-length event subsequences and the fixed-length parameter subsequences alike, whereby those subsequences each have a same length.
The motivation to combine (Jianwu Xu), (Kraus), and (Ali) as to Claim 11 is as recited by (Ali), namely that governing the extraction by a single user-defined parameter k yields feature vectors of a determinate and mutually commensurable length, since (Ali) teaches that the resulting vector has a "length |Σ|ᵏ, which will contain the exact number of occurrences of each possible k-mer" (Ali, p. 157, § 3.1, ¶ 2), and that "the kernel value for these vectors is simply the dot product of A and B" (Ali, p. 157, § 3.1, ¶ 3). One of ordinary skill in the art would have recognized that a dot product is defined only between vectors of equal length, such that applying a common value of k to the event subsequences and the parameter subsequences is necessary to permit the comparison operation that (Ali) teaches, yielding the predictable result of subsequences of a same length.
As to dependent Claim 12, the claim recites:
The method according to claim 10, the determining the respective fixed-length embedding of the respective event sequence further comprising:
determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences;
determining a second frequency vector based on the plurality of fixed-length event subsequences, the second frequency vector including values that each indicate a number of occurrences of a respective possible event subsequence in the plurality of fixed-length event subsequences; and
determining the respective fixed-length embedding of the respective event sequence based on the first frequency vector and the second frequency vector.
Regarding the limitation "The method according to claim 10, the determining the respective fixed-length embedding of the respective event sequence further comprising:", Claim 12 depends from Claim 10, which in turn depends from Claim 9 and Claim 1, and Claim 12 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1, Claim 9, and Claim 10 over (Jianwu Xu) in view of (Kraus), further in view of (Panpan Xu), and further in view of (Ali).
Regarding Claim 12, the combination of (Jianwu Xu) and (Kraus) does not teach:
"determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences,"
"determining a second frequency vector based on the plurality of fixed-length event subsequences, the second frequency vector including values that each indicate a number of occurrences of a respective possible event subsequence in the plurality of fixed-length event subsequences,"
"determining the respective fixed-length embedding of the respective event sequence based on the first frequency vector and the second frequency vector."
In the same field of endeavor, (Ali) teaches each of the three limitations added by Claim 12, as set forth separately below.
(Ali) teaches "determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences." Specifically, (Ali) discloses that "[g]iven X, k, and Σ, we have to design a frequency feature vector Φₖ(X) of length |Σ|ᵏ, which will contain the exact number of occurrences of each possible k-mer in X" (Ali, p. 157, § 3.1, ¶ 2). The recited vector Φₖ(X) is a frequency vector because its stored values are occurrence counts, as (Ali) further confirms that the algorithm "takes the feature vectors (containing a count of each k-mers) as input" (Ali, p. 157, § 3.1, ¶ 3). The recited vector includes a value for each possible subsequence rather than only for those subsequences actually present, since (Ali) discloses that "since we do not know the total unique k-mers in all spike sequences, we need to consider all possible pairs of k-mers to design a general purpose feature vector representation" (Ali, p. 156, § 3.1, ¶ 2), and that the vector has length |Σ|ᵏ, which is the count of all possible k-length subsequences over the alphabet Σ. Applied to the plurality of fixed-length parameter subsequences of Claim 10, this disclosure teaches a first frequency vector whose values each indicate the number of occurrences of a respective possible parameter subsequence among those parameter subsequences.
(Ali) teaches "determining a second frequency vector based on the plurality of fixed-length event subsequences, the second frequency vector including values that each indicate a number of occurrences of a respective possible event subsequence in the plurality of fixed-length event subsequences." Specifically, (Ali) discloses that the frequency feature vector construction is applied to each sequence of a dataset independently and that a separate vector is thereby obtained for each such sequence, since "[t]he process of kernel value computation is repeated for each pair of sequences" (Ali, p. 157, § 3.1, ¶ 3), which repetition presupposes that a respective feature vector has been computed for each sequence. (Ali) further discloses the pairwise use of two such separately computed vectors, stating that "[g]iven two feature vectors A and B, the kernel value for these vectors is simply the dot product of A and B. For example, given a k-mer, if the frequency of that k-mer in A is 2 and B is 3" (Ali, p. 157, § 3.1, ¶ 3), thereby teaching that the same occurrence-counting construction yields a second frequency vector indexed by the same set of possible subsequences. Applied to the plurality of fixed-length event subsequences of Claim 10, this disclosure teaches a second frequency vector whose values each indicate the number of occurrences of a respective possible event subsequence among those event subsequences.
(Ali) teaches "determining the respective fixed-length embedding of the respective event sequence based on the first frequency vector and the second frequency vector." Specifically, (Ali) discloses that the frequency feature vector is itself the fixed-length representation of the sequence from which it was derived, since the vector has a "length |Σ|ᵏ" (Ali, p. 157, § 3.1, ¶ 2) that is fixed by the alphabet Σ and the parameter k and is therefore independent of the length N of the underlying sequence, and since (Ali) frames the entire operation as being "[f]or mapping protein sequences to fixed-length vectors" (Ali, p. 156, § 3.1, ¶ 1). (Ali) further discloses that the frequency vectors so determined are carried forward as the representation used in the downstream analysis, stating that "the reduced dimensional principal components-based feature vector representation is given as input to the classical machine learning models" (Ali, p. 156, § 3, ¶ 1). Applied to the first and second frequency vectors recited above, this disclosure teaches determining the respective fixed-length embedding of the respective event sequence based on both frequency vectors.
(Jianwu Xu), (Kraus), and (Ali) are analogous to the claimed invention for the reasons set forth above in the rejection of Claim 10, namely that (Jianwu Xu) and (Kraus) are from the same field of endeavor of analyzing time-stamped event records generated by a monitored computing or industrial system, and (Ali) is reasonably pertinent to the particular problem with which the inventors were concerned, namely the transformation of variable-length sequences of discrete symbols into fixed-length vector representations. See MPEP § 2141.01(a). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence and parameter sequence construction of (Jianwu Xu) and (Kraus) with the frequency feature vector of (Ali).
The motivation to combine (Jianwu Xu), (Kraus), and (Ali) is as recited by (Ali), namely that a frequency vector indexed by every possible subsequence yields a representation whose length is fixed by the alphabet and the parameter k rather than by the length of the underlying sequence, since (Ali) teaches that the vector is designed "[f]or mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1). One of ordinary skill in the art would have been motivated to apply this construction to the event subsequences and parameter subsequences of (Jianwu Xu) and (Kraus) because Claim 1 as taught by that combination requires embeddings of a fixed length notwithstanding that the underlying event sequences have variable lengths (Kraus, col. 25, lines 6-9), and because occurrence counting over the set of all possible subsequences supplies precisely such a length-invariant representation while retaining the ordering information that (Jianwu Xu) identifies as otherwise lost, stating that "information regarding the sequential or temporal order of the event types is often not well represented" (Jianwu Xu, col. 12, lines 24-27).
As to dependent Claim 14, the claim recites:
The method according to claim 12, wherein:
the first frequency vector has a length equal to a total number possible parameter subsequences; and
the second frequency vector has a length equal to a total number possible event subsequences.
Regarding Claim 14, the combination of (Jianwu Xu) and (Kraus) does not teach:
"the first frequency vector has a length equal to a total number possible parameter subsequences"
"the second frequency vector has a length equal to a total number possible event subsequences."
In the same field of endeavor, (Ali) teaches each of the two limitations added by Claim 14, as set forth separately below.
(Ali) teaches "the first frequency vector has a length equal to a total number possible parameter subsequences." Specifically, (Ali) discloses that "[g]iven X, k, and Σ, we have to design a frequency feature vector Φₖ(X) of length |Σ|ᵏ, which will contain the exact number of occurrences of each possible k-mer in X" (Ali, p. 157, § 3.1, ¶ 2). The quantity |Σ|ᵏ is by definition the total number of distinct subsequences of length k that can be formed over the alphabet Σ, where (Ali) defines Σ as "an alphabet Σ (a finite set of symbols)" (Ali, p. 157, § 3.1, ¶ 2). (Ali) further confirms that the vector is dimensioned to the enumeration of all possible subsequences rather than to those actually present in any given sequence, disclosing that "since we do not know the total unique k-mers in all spike sequences, we need to consider all possible pairs of k-mers to design a general purpose feature vector representation for spike sequences in a given dataset" (Ali, p. 156, § 3.1, ¶ 2), and characterizing the consequence of that choice as "the huge dimensionality of the feature vector |Σ|ᵏ that can make kernel computation very expensive" (Ali, p. 157, § 3.1, ¶ 2). Applied to the plurality of fixed-length parameter subsequences of Claim 10, this disclosure teaches a first frequency vector whose length is equal to the total number of possible parameter subsequences.
(Ali) teaches "the second frequency vector has a length equal to a total number possible event subsequences." Specifically, (Ali) discloses that the same construction of length |Σ|ᵏ is applied to yield a respective feature vector for each sequence to be represented, since the operation is defined generally for any sequence "X ∈ Σ" over the alphabet (Ali, p. 157, § 3.1, ¶ 2), and since "[t]he process of kernel value computation is repeated for each pair of sequences" (Ali, p. 157, § 3.1, ¶ 3), which repetition presupposes that a feature vector of that same length has been computed for each sequence. (Ali) further discloses that two such vectors are of equal and commensurable length, stating that "[g]iven two feature vectors A and B, the kernel value for these vectors is simply the dot product of A and B. For example, given a k-mer, if the frequency of that k-mer in A is 2 and B is 3, its contribution towards the kernel value of A and B is simply 2·3" (Ali, p. 157, § 3.1, ¶ 3), which per-k-mer correspondence between vectors A and B is possible only because both vectors are indexed by the same enumeration of all possible k-mers. Applied to the plurality of fixed-length event subsequences of Claim 10, this disclosure teaches a second frequency vector whose length is equal to the total number of possible event subsequences.
(Jianwu Xu), (Kraus), and (Ali) are analogous to the claimed invention as all three are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13), (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-10), and (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the first and second frequency vectors of (Jianwu Xu) and (Kraus) with the |Σ|ᵏ vector dimensioning of (Ali), such that the first frequency vector has a length equal to the total number of possible parameter subsequences and the second frequency vector has a length equal to the total number of possible event subsequences.
The motivation to combine (Jianwu Xu), (Kraus), and (Ali) is as recited by (Ali), namely that dimensioning the vector to the enumeration of all possible subsequences yields a representation that is valid across every sequence of the dataset irrespective of which subsequences any individual sequence happens to contain, since (Ali) teaches that "since we do not know the total unique k-mers in all spike sequences, we need to consider all possible pairs of k-mers to design a general purpose feature vector representation for spike sequences in a given dataset" (Ali, p. 156, § 3.1, ¶ 2). One of ordinary skill in the art would have been motivated to adopt this dimensioning in the method of (Jianwu Xu) and (Kraus) because Claim 1 as taught by that combination requires that a cluster analysis be performed across the plurality of embeddings (Panpan Xu, col. 19, lines 25-33), and a cluster analysis requires that every embedding occupy the same vector space, which the |Σ|ᵏ dimensioning of (Ali) guarantees.
Claims 13 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, further in view of Ali et al. (Ali), Non-Patent Literature, "A k-mer Based Approach for SARS-CoV-2 Variant Identification," International Symposium on Bioinformatics Research and Applications (ISBRA), 2021, pp. 153–164, cited in the IDS filed 03/31/2023, and further in view of Zhuang et al. (Zhuang), US 2020/0050941 A1.
As to dependent Claim 13, the claim recites:
The method according to claim 12, the determining the respective fixed-length embedding of the respective event sequence further comprising:
determining the respective fixed-length embedding of the respective event sequence as a concatenation of the first frequency vector and the second frequency vector.
Regarding the limitation "The method according to claim 12, the determining the respective fixed-length embedding of the respective event sequence further comprising:", Claim 13 depends from Claim 12, which in turn depends from Claim 10, Claim 9, and Claim 1, and Claim 13 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1, Claim 9, Claim 10, and Claim 12 over (Jianwu Xu) in view of (Kraus), further in view of (Panpan Xu), and further in view of (Ali).
Regarding Claim 13, the combination of (Jianwu Xu), (Kraus), and (Ali) teaches the first frequency vector and the second frequency vector as set forth above in the rejection of Claim 12. (Ali) further teaches something related to "determining the respective fixed-length embedding of the respective event sequence as a concatenation of the first frequency vector and the second frequency vector." However, the combination of (Jianwu Xu), (Kraus), and (Ali) does not teach "determining the respective fixed-length embedding of the respective event sequence as a concatenation of the first frequency vector and the second frequency vector."
In the same field of endeavor, (Zhuang) teaches this limitation.
(Zhuang) teaches determining an embedding as a concatenation of two constituent vectors. Specifically, (Zhuang) discloses that "the attribute network is operatively coupled to the sequence network via a fusion network that comprises an input concatenation layer which is configured to concatenate an output of the attribute vector output layer with an output of the sequence network module" (Zhuang, ¶ [0017]). (Zhuang) further discloses the same operation with reference to FIG. 9, stating that "an attribute network 902 and a sequence network 904 are coupled to a fusion network 906 comprising a concatenation layer 908 and a fully-connected layer 910," where "the output V of the M'-th layer of the attribute network 902 is coupled to the concatenation layer 908 via a connection 912, and the hidden state of the sequence network 904 after processing of the last time step is coupled to the concatenation layer 908 via a connection 914" (Zhuang, ¶ [0069]). (Zhuang) further discloses the operation in the alternative arrangement of FIG. 8, stating that "[t]he coupling may be effected by concatenating the hidden state of the sequence network 802 after processing of the last time step with the attribute data," and identifying the operator expressly as "the concatenation operator" (Zhuang, ¶ [0068]).
(Zhuang) teaches that the vector so formed by concatenation is the fixed-length embedding. Specifically, (Zhuang) discloses that "the fixed-length feature representation of input attributed sequence data may comprise an output vector of the fully-connected feedforward neural network layer" that operates on the concatenated inputs (Zhuang, ¶ [0017]), and that "the number of features in the resulting embedding is equal to the size of the output of the nonlinear function module, and in particular may be equal to the sum of the first and second predetermined numbers, i.e., the combined count of units in the attribute vector output layer and hidden units in the LSTM network" (Zhuang, ¶ [0018]). (Zhuang) further discloses that the number of units in the concatenation layer "may be equal to a sum of the first predetermined number of attribute features and the second predetermined number of sequence features" (Zhuang, ¶ [0017]), thereby teaching that the length of the concatenated embedding is the sum of the lengths of the two constituent vectors and is fixed by those two predetermined numbers.
(Jianwu Xu), (Kraus), (Ali), and (Zhuang) are analogous to the claimed invention as all four are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13), (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-10), (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1), and (Zhuang) is directed to "variable-length sequences of categorical items" for which "the goal is to transform a variable-length sequence into a fixed-length feature representation" (Zhuang, ¶ [0002]). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the first and second frequency vectors of (Jianwu Xu), (Kraus), and (Ali) with the concatenation layer of (Zhuang), such that the respective fixed-length embedding of the respective event sequence is determined as a concatenation of those two frequency vectors.
The motivation to combine (Jianwu Xu), (Kraus), (Ali), and (Zhuang) is as recited by (Zhuang), namely that concatenating the two constituent representations preserves the information carried by each while permitting the interdependencies between them to be captured, since (Zhuang) teaches a "nonlinear function over the concatenation to capture the dependencies between attributes and sequences" (Zhuang, ¶ [0069]), whereas the dot product taught by (Ali) collapses two vectors into a single scalar and thereby discards the constituent values (Ali, p. 157, § 3.1, ¶ 3).
As to dependent Claim 15, the claim recites:
The method according to claim 10, the determining the plurality of fixed-length embeddings comprising:
determining a kernel matrix having values indicating a similarity between each possible combination of two event sequences in the plurality of event sequences; and
determining the respective fixed-length embedding of each respective event sequence in the plurality of event sequences by:
determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences;
determining a kernel vector for the respective event sequence the based on the kernel matrix; and
determining the respective fixed-length embedding of the respective event sequence based on the kernel vector and the first frequency vector.
Claim 15 depends from Claim 10, which in turn depends from Claim 9 and Claim 1, and Claim 15 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1, Claim 9, and Claim 10 over (Jianwu Xu) in view of (Kraus), further in view of (Panpan Xu), and further in view of (Ali).
Regarding Claim 15, the combination of (Jianwu Xu) and (Kraus) does not teach:
"determining a kernel matrix having values indicating a similarity between each possible combination of two event sequences in the plurality of event sequences,"
"determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences,"
"determining a kernel vector for the respective event sequence the based on the kernel matrix."
In the same field of endeavor, (Ali) teaches "determining a kernel matrix having values indicating a similarity between each possible combination of two event sequences in the plurality of event sequences." Specifically, (Ali) discloses that "[t]he process of kernel value computation is repeated for each pair of sequences and hence we get a (symmetric) matrix (kernel matrix) containing a similarity score between each pair of sequences" (Ali, p. 157, § 3.1, ¶ 3), and states this operation as the object of the disclosed approach, namely that "[g]iven a set of spike sequences, our goal is to find the similarity score between each pair of sequences (kernel matrix)" (Ali, p. 156, § 3, ¶ 1). Because the computation is repeated for each pair and the resulting matrix is symmetric, the matrix so assembled contains a value for each possible combination of two sequences in the plurality.
(Ali) teaches "determining a first frequency vector based on the plurality of fixed-length parameter subsequences, the first frequency vector including values that each indicate a number of occurrences of a respective possible parameter subsequence in the plurality of fixed-length parameter subsequences." Specifically, (Ali) discloses that "[g]iven X, k, and Σ, we have to design a frequency feature vector Φₖ(X) of length |Σ|ᵏ, which will contain the exact number of occurrences of each possible k-mer in X" (Ali, p. 157, § 3.1, ¶ 2), and confirms the occurrence-count construction by describing the algorithm as one that "takes the feature vectors (containing a count of each k-mers) as input" (Ali, p. 157, § 3.1, ¶ 3). The vector so designed contains one value per possible subsequence, each value being the number of occurrences of that subsequence among the subsequences extracted.
(Ali) teaches "determining a kernel vector for the respective event sequence the based on the kernel matrix." Specifically, (Ali) discloses, under the heading "3.2 Kernel PCA," that "[d]ue to a high-dimensional kernel matrix, we use Kernel PCA (K-PCA) to select a subset of principal components. These extracted principal components corresponding to each spike sequence act as the feature vector representations for the spike sequences (we selected 50 principal components for our experiments)" (Ali, p. 157, § 3.2), and that "[t]he resultant kernel matrix is given as input to the kernel PCA method for dimensionality reduction" (Ali, p. 156, § 3, ¶ 1). The per-sequence vector of principal components so extracted is a vector determined for the respective sequence based on the kernel matrix.
The combination of (Jianwu Xu), (Kraus), and (Ali), however, does not teach "determining the respective fixed-length embedding of the respective event sequence based on the kernel vector and the first frequency vector."
In the same field of endeavor, (Zhuang) teaches this limitation. Specifically, (Zhuang) discloses forming a fixed-length embedding from two separately computed constituent vectors, namely a "fusion network that comprises an input concatenation layer which is configured to concatenate an output of the attribute vector output layer with an output of the sequence network module, and a nonlinear function module that is configured to learn a nonlinear function of the concatenated inputs," and that "the fixed-length feature representation of input attributed sequence data may comprise an output vector of the fully-connected feedforward neural network layer" so configured (Zhuang, ¶ [0017]). (Zhuang) further discloses the same arrangement with reference to FIG. 9, stating that "an attribute network 902 and a sequence network 904 are coupled to a fusion network 906 comprising a concatenation layer 908 and a fully-connected layer 910 implementing a nonlinear function over the concatenation" (Zhuang, ¶ [0069]), and that the length of the resulting embedding is fixed by the two constituent lengths, since "the number of features in the resulting embedding . . . may be equal to the sum of the first and second predetermined numbers" (Zhuang, ¶ [0018]).
(Jianwu Xu), (Kraus), (Ali), and (Zhuang) are analogous to the claimed invention as all four are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13), (Kraus) is directed to "a system's event log 302, which documents a time-series of events 204 that occurred in the system 130" (Kraus, col. 28, lines 8-10), (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1), and (Zhuang) is directed to "variable-length sequences of categorical items" for which "the goal is to transform a variable-length sequence into a fixed-length feature representation" (Zhuang, ¶ [0002]). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the kernel matrix, frequency vector, and kernel vector of (Ali) with the fusion arrangement of (Zhuang).
The motivation to combine (Jianwu Xu), (Kraus), (Ali), and (Zhuang) is as recited by (Zhuang), namely that forming the embedding from two constituent vectors preserves the information carried by each while permitting the interdependencies between them to be captured, since (Zhuang) teaches a "nonlinear function over the concatenation to capture the dependencies between attributes and sequences" (Zhuang, ¶ [0069]), whereas (Ali) discards the frequency vectors once the kernel matrix has been computed, retaining only "a subset of principal components" (Ali, p. 157, § 3.2).
Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, further in view of Ali et al. (Ali), Non-Patent Literature, "A k-mer Based Approach for SARS-CoV-2 Variant Identification," International Symposium on Bioinformatics Research and Applications (ISBRA), 2021, pp. 153–164, cited in the IDS filed 03/31/2023, further in view of Zhuang et al. (Zhuang), US 2020/0050941 A1, and further in view of Leslie et al. (Leslie), Non-Patent Literature, "Mismatch String Kernels for SVM Protein Classification," Advances in Neural Information Processing Systems (NeurIPS), 2003, pp. 1441–1448, cited in the IDS filed 03/31/2023.
As to dependent Claim 16, the claim recites:
The method according to claim 15, the determining the kernel matrix further comprising:
for each respective combination of two event sequences in the plurality of event sequences, determining a respective value in the kernel matrix for the respective combination of two event sequences by:
determining a respective second frequency vector, the respective second frequency vector including values that each indicate a number of occurrences, with a predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a first event sequence in the respective combination of two event sequences;
determining a respective third frequency vector, the respective third frequency vector including values that each indicate a number of occurrences, with the predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a second event sequence in the respective combination of two event sequences; and
determining the respective value in the kernel matrix for the respective combination of two event sequences as a dot product of the respective second frequency vector and the respective third frequency vector.
Claim 16 depends from Claim 15, which in turn depends from Claim 10, Claim 9, and Claim 1, and Claim 16 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1, Claim 9, Claim 10, and Claim 15 over (Jianwu Xu) in view of (Kraus), further in view of (Panpan Xu), further in view of (Ali), and further in view of (Zhuang).
Regarding Claim 16, (Jianwu Xu) teaches something related to "for each respective combination of two event sequences in the plurality of event sequences, determining a respective value in the kernel matrix for the respective combination of two event sequences." However, (Jianwu Xu) does not teach this limitation.
In the same field of endeavor, (Ali) teaches "for each respective combination of two event sequences in the plurality of event sequences, determining a respective value in the kernel matrix for the respective combination of two event sequences." Specifically, (Ali) discloses that "[t]he process of kernel value computation is repeated for each pair of sequences and hence we get a (symmetric) matrix (kernel matrix) containing a similarity score between each pair of sequences" (Ali, p. 157, § 3.1, ¶ 3), and states this operation as the object of the disclosed approach, namely that "[g]iven a set of spike sequences, our goal is to find the similarity score between each pair of sequences (kernel matrix)" (Ali, p. 156, § 3, ¶ 1). Because the computation is repeated for each pair and the resulting matrix is symmetric, a respective value is determined for each respective combination of two sequences in the plurality.
(Ali) teaches "determining the respective value in the kernel matrix for the respective combination of two event sequences as a dot product of the respective second frequency vector and the respective third frequency vector." Specifically, (Ali) discloses that "[g]iven two feature vectors A and B, the kernel value for these vectors is simply the dot product of A and B. For example, given a k-mer, if the frequency of that k-mer in A is 2 and B is 3, its contribution towards the kernel value of A and B is simply 2·3" (Ali, p. 157, § 3.1, ¶ 3). (Ali) further discloses that the two vectors so combined are frequency vectors, describing the algorithm as one that "takes the feature vectors (containing a count of each k-mers) as input and returns a real-valued similarity score between each pair of vectors" (Ali, p. 157, § 3.1, ¶ 3).
(Jianwu Xu) and (Ali) are analogous to the claimed invention as both are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence analysis of (Jianwu Xu) with the pairwise kernel matrix and dot product computation of (Ali). The motivation to combine (Jianwu Xu) and (Ali) is as recited by (Ali), namely that computing a similarity score between each pair of sequences yields a general purpose representation applicable across an entire dataset, since (Ali) teaches that "we need to consider all possible pairs of k-mers to design a general purpose feature vector representation for spike sequences in a given dataset" (Ali, p. 156, § 3.1, ¶ 2), which serves the objective of (Jianwu Xu) of grouping related event patterns so as to "help pinpoint the causes of system faults and failures" (Jianwu Xu, col. 2, lines 46-48).
The combination of (Jianwu Xu) and (Ali), however, does not teach:
"determining a respective second frequency vector, the respective second frequency vector including values that each indicate a number of occurrences, with a predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a first event sequence in the respective combination of two event sequences,"
"determining a respective third frequency vector, the respective third frequency vector including values that each indicate a number of occurrences, with the predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a second event sequence in the respective combination of two event sequences."
In the same field of endeavor, (Leslie) teaches "determining a respective second frequency vector, the respective second frequency vector including values that each indicate a number of occurrences, with a predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a first event sequence in the respective combination of two event sequences." Specifically, (Leslie) discloses that "[t]he (k,m)-mismatch kernel is based on a feature map to a vector space indexed by all possible subsequences of amino acids of a fixed length k; each instance of a fixed k-length subsequence in an input sequence contributes to all feature coordinates differing from it by at most m mismatches" (Leslie, p. 1442, § 1, ¶ 4). (Leslie) further defines the mismatch neighborhood by reference to a predetermined number m, stating that "[f]or a fixed k-mer α = a₁a₂…aₖ, with each aᵢ a character in A, the (k,m)-neighborhood generated by α is the set of all k-length sequences β from A that differ from α by at most m mismatches" (Leslie, p. 1442, § 2.1, ¶ 2). (Leslie) further discloses that the value stored at each coordinate of the resulting vector is the mismatch-tolerant occurrence count, stating that "the β-coordinate of φ(k,m)(x) is just a count of all instances of the k-mer β occurring with up to m mismatches in x" (Leslie, p. 1443, § 2.1, ¶ 1). (Leslie) further discloses that the vector is formed from the fixed-length subsequences extracted from the input sequence, stating that "[f]or a sequence x of any length, we extend the map additively by summing the feature vectors for all the k-mers in x" (Leslie, p. 1442, § 2.1, ¶ 3 – p. 1443, ¶ 1).
(Leslie) teaches "determining a respective third frequency vector, the respective third frequency vector including values that each indicate a number of occurrences, with the predetermined number of mismatches or less, of a respective possible event subsequence in the respective plurality of fixed-length event subsequences of a second event sequence in the respective combination of two event sequences." Specifically, (Leslie) discloses that the identical feature map, governed by the same predetermined values of k and m, is applied to the second sequence of the pair, since the kernel is defined between two such vectors as "K(k,m)(x,y) = ⟨φ(k,m)(x), φ(k,m)(y)⟩" (Leslie, p. 1443, § 2.1, ¶ 1), where φ(k,m)(y) is the vector produced by applying that map to the second sequence y. (Leslie) further confirms that both vectors are indexed by the same enumeration of possible subsequences, disclosing that the feature map maps "from the space of all finite sequences from an alphabet A of size |A| = l to the lᵏ-dimensional vector space indexed by the set of k-length subsequences ('k-mers') from A" (Leslie, p. 1442, § 2.1, ¶ 1).
(Jianwu Xu), (Ali), and (Leslie) are analogous to the claimed invention as all three are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13), (Ali) is directed to "mapping protein sequences to fixed-length vectors" (Ali, p. 156, § 3.1, ¶ 1), and (Leslie) is directed to "represent[ing] protein sequences as vectors in a high-dimensional feature space via a string-based feature map" (Leslie, p. 1442, § 2, ¶ 1). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the kernel matrix and dot product computation of (Jianwu Xu) and (Ali) with the mismatch-tolerant feature map of (Leslie), such that the second and third frequency vectors whose dot product yields each kernel matrix value are populated with occurrence counts taken with a predetermined number of mismatches or less.
The motivation to combine (Jianwu Xu), (Ali), and (Leslie) is as recited by (Leslie), namely that tolerating a bounded number of mismatches permits the kernel to recognize similarity between sequences that are only remotely related, since (Leslie) teaches that its object is "the more difficult problem of remote homology detection, where we want our classifier to detect (as positives) test sequences that are only remotely related to the positive training sequences" (Leslie, p. 1441, § 1, ¶ 1), and that the mismatch tolerance is what supplies that capability, since "the mismatch kernel adds the biologically important idea of mismatching to the computationally simpler spectrum kernel" (Leslie, p. 1442, § 1, ¶ 4). (Leslie) further teaches that the technique is not confined to biological sequences, stating that "the mismatch kernel does not depend on any generative model and could potentially be used in other sequence-based classification problems" (Leslie, p. 1442, § 1, ¶ 5). One of ordinary skill in the art would have been motivated to apply the mismatch tolerance of (Leslie) to the exact-count frequency vectors of (Ali) because (Ali) already contemplates mismatch counting as a measure of sequence distance (Ali, p. 157, § 3.1, ¶ 2), and because event sequences drawn from a monitored system rarely repeat exactly, such that a kernel requiring exact subsequence matches would fail to group event sequences differing only in isolated events.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, further in view of Ali et al. (Ali), Non-Patent Literature, "A k-mer Based Approach for SARS-CoV-2 Variant Identification," International Symposium on Bioinformatics Research and Applications (ISBRA), 2021, pp. 153–164, cited in the IDS filed 03/31/2023, further in view of Zhuang et al. (Zhuang), US 2020/0050941 A1, and further in view of Hoffmann (Hoffmann), Non-Patent Literature, "Kernel PCA for Novelty Detection," Pattern Recognition, vol. 40, no. 3, pp. 863–874, 2007, cited in the IDS filed 03/31/2023.
As to dependent Claim 17, the claim recites:
The method according to claim 15, the determining the kernel vector further comprising:
determining the kernel vector for the respective event sequence by applying kernel principal component analysis to the kernel matrix.
Claim 17 depends from Claim 15, which in turn depends from Claim 10, Claim 9, and Claim 1, and Claim 17 incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejections of Claim 1, Claim 9, Claim 10, and Claim 15 over (Jianwu Xu) in view of (Kraus), further in view of (Panpan Xu), further in view of (Ali), and further in view of (Zhuang).
Regarding Claim 17, (Jianwu Xu) teaches something related to "determining the kernel vector for the respective event sequence by applying kernel principal component analysis to the kernel matrix." However, (Jianwu Xu) does not teach this limitation.
In the same field of endeavor, (Ali) teaches this limitation. Specifically, (Ali) discloses, under the heading "3.2 Kernel PCA," that "[d]ue to a high-dimensional kernel matrix, we use Kernel PCA (K-PCA) to select a subset of principal components. These extracted principal components corresponding to each spike sequence act as the feature vector representations for the spike sequences (we selected 50 principal components for our experiments)" (Ali, p. 157, § 3.2). (Ali) further discloses that the kernel matrix is the input to that operation, stating that "[t]he resultant kernel matrix is given as input to the kernel PCA method for dimensionality reduction" (Ali, p. 156, § 3, ¶ 1), and that the vector so produced is carried forward as the representation of the respective sequence, since "[t]he reduced dimensional principal components-based feature vector representation is given as input to the classical machine learning models" (Ali, p. 156, § 3, ¶ 1). Because the extracted principal components correspond to each individual sequence and are obtained by applying kernel principal component analysis to the kernel matrix, these disclosures teach determining the kernel vector for the respective event sequence by applying kernel principal component analysis to the kernel matrix.
(Jianwu Xu) and (Ali) are analogous to the claimed invention as both are from the same field of endeavor of transforming sequences of discrete data elements into vector representations for machine learning analysis, since (Jianwu Xu) is directed to event record logs generated by "Information and Communication Technology systems and manufacturing plant systems" (Jianwu Xu, col. 1, lines 11-13) and (Ali) is directed to "mapping protein sequences to fixed-length vectors" while preserving "order information" (Ali, p. 156, § 3.1, ¶ 1). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence analysis of (Jianwu Xu) with the kernel principal component analysis of (Ali), such that the kernel vector for each respective event sequence is determined by applying kernel principal component analysis to the kernel matrix.
The motivation to combine (Jianwu Xu) and (Ali) is as recited by (Ali), namely that applying kernel principal component analysis to the kernel matrix reduces the dimensionality of an otherwise unmanageably large representation, since (Ali) teaches that "[d]ue to a high-dimensional kernel matrix, we use Kernel PCA (K-PCA) to select a subset of principal components" (Ali, p. 157, § 3.2) and identifies the underlying problem as "the huge dimensionality of the feature vector |Σ|ᵏ that can make kernel computation very expensive" (Ali, p. 157, § 3.1, ¶ 2). One of ordinary skill in the art would have been motivated to apply this dimensionality reduction in the method of (Jianwu Xu) because (Jianwu Xu) likewise operates on a reduced-dimension latent representation for reasons of computational tractability, selecting "any value between 100 and 300" as the number of dimensions (Jianwu Xu, col. 11, lines 1-4), and because (Jianwu Xu) expressly reduces its event records to identifiers "for ease of computation and manipulation in the subsequent steps" (Jianwu Xu, col. 10, lines 46-49).
The combination of (Jianwu Xu) and (Ali), however, does not detail the operation by which kernel principal component analysis derives a per-sequence coefficient vector from the kernel matrix. In the same field of endeavor, (Hoffmann) teaches that operation. Specifically, (Hoffmann) discloses that "[k]ernel principal component analysis (kernel PCA) is a non-linear extension of PCA. . . . Training data are mapped into an infinite-dimensional feature space. In this space, kernel PCA extracts the principal components of the data distribution" (Hoffmann, p. 863, Abstract). (Hoffmann) further discloses that the principal components are expressed as coefficient vectors over the data points, stating that "[i]n kernel PCA, an eigenvector V of the covariance matrix in F is a linear combination of points Φ(xᵢ), V = Σᵢ αᵢ Φ̃(xᵢ)" (Hoffmann, p. 864, § 2). (Hoffmann) further discloses that the coefficient vector α is obtained from the kernel matrix itself, stating that "[t]he αᵢ are the components of a vector α. It turns out that this vector is an eigenvector of the matrix K̃ᵢⱼ = (Φ̃(xᵢ)·Φ̃(xⱼ))," and that "[t]o compute K̃, we substitute Φ̃ according to Eq. (3). This substitution gives K̃ᵢⱼ as a function of the kernel matrix Kᵢⱼ = k(xᵢ,xⱼ)" (Hoffmann, p. 864, § 2).
(Jianwu Xu), (Ali), and (Hoffmann) are analogous to the claimed invention as all three are from the same field of endeavor of transforming data into reduced-dimension vector representations for machine learning analysis, since (Jianwu Xu) is directed to the "latent representation of event record types" (Jianwu Xu, col. 11, lines 1-4), (Ali) is directed to "mapping protein sequences to fixed-length vectors" (Ali, p. 156, § 3.1, ¶ 1), and (Hoffmann) is directed to a method by which "kernel PCA extracts the principal components of the data distribution" from data mapped into a feature space (Hoffmann, p. 863, Abstract). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the kernel principal component analysis of (Jianwu Xu) and (Ali) with the eigenvector computation of (Hoffmann), such that the kernel vector determined for each respective event sequence is the coefficient vector obtained as an eigenvector of the centered kernel matrix.
The motivation to combine (Jianwu Xu), (Ali), and (Hoffmann) is as recited by (Hoffmann), namely that the kernel formulation permits the principal components to be computed without ever forming the high-dimensional feature vectors themselves, since (Hoffmann) teaches that "[t]he trick herein is that the PCA can be computed such that the vectors Φ(xᵢ) appear only within scalar products. Thus, mapping (1) can be omitted. Instead, we only work with a kernel function k(x,y), which replaces the scalar product" (Hoffmann, p. 864, § 2). One of ordinary skill in the art would have been motivated to adopt this computation because (Ali) identifies the expense of operating directly on the |Σ|ᵏ-dimensional feature vectors as the very problem to be avoided, stating that "in the so-called kernel trick, kernel values are directly evaluated instead of comparing indices" (Ali, p. 157, § 3.1, ¶ 2), and (Hoffmann) supplies the corresponding computation for the principal component extraction step.
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Xu et al. (Jianwu Xu), US 11,294,754 B2, in view of Kraus et al. (Kraus), US 11,106,789 B2, further in view of Xu et al. (Panpan Xu), US 11,074,276 B2, and further in view of Aggarwal et al. (Aggarwal), US 2022/0019888 A1.
As to dependent Claim 20, the claim recites:
The method according to claim 1 further comprising:
predicting, with the processor, a possible future event based on a partial event sequence and the plurality of clusters of event sequences.
Claim 20 depends from Claim 1 and incorporates all of the limitations thereof. Those limitations are rejected under the same rationale set forth above in the rejection of Claim 1 over (Jianwu Xu) in view of (Kraus) and further in view of (Panpan Xu).
Regarding Claim 20, (Jianwu Xu) teaches something related to "predicting, with the processor, a possible future event based on a partial event sequence and the plurality of clusters of event sequences." However, (Jianwu Xu) does not teach this limitation.
In the same field of endeavor, (Aggarwal) teaches this limitation.
(Aggarwal) teaches "predicting, with the processor, a possible future event." Specifically, (Aggarwal) discloses that the trained model is used, "given a sequence of events as input, [to] predict a next event to occur after the last event in the sequence," and that "[a]s part of predicting the next event, the unified model is trained to predict an event type for the next event and a time of occurrence for the next event" (Aggarwal, ¶ [0004]). (Aggarwal) further discloses that the operation is performed by a processor-implemented system, stating that "a computer system receives, using a neural network, a sequence of events and time of occurrence associated with each event," and that "[t]he computer system, using the neural network and based upon the vector representations, predicts a next event to occur after the sequence of events" (Aggarwal, ¶ [0006]).
(Aggarwal) teaches that the prediction is made "based on a partial event sequence." Specifically, (Aggarwal) discloses that the input to the prediction is a sequence of events that terminates before the event to be predicted, since the model predicts "a next event to occur after the last event in the sequence" (Aggarwal, ¶ [0004]), and since the clustering information is output "for the sequence of events not including the next event" (Aggarwal, ¶ [0006]). The sequence supplied as input is therefore a partial event sequence relative to the completed sequence that would include the predicted event.
(Aggarwal) teaches that the prediction is made "based on . . . the plurality of clusters of event sequences." Specifically, (Aggarwal) discloses that the prediction and the clustering are performed jointly by a single model rather than as independent operations, stating that the disclosure describes "using a single unified machine learning model (e.g., a neural network) for performing both supervised event predictions and unsupervised time-varying clustering for a sequence of events," and that the model is trained "for both predicting next events and performing dynamic clustering for a sequence of events" (Aggarwal, ¶ [0004]). (Aggarwal) further discloses that the two operations are jointly optimized, stating that "supervised predictions (e.g. event type and time of occurrence predictions) and unsupervised dynamic clustering are jointly performed in a neural network using a unified model," which "enables systems to predict both an event type and associated time of occurrence for the next event in a sequence of events and to indicate an evolution of user behavior in the sequence of events through clustering over the sequence of events" (Aggarwal, ¶ [0005]). Because a single model produces both outputs from a common set of learned representations, the predicted next event is determined based in part on the clustering.
(Jianwu Xu) and (Aggarwal) are analogous to the claimed invention as both are from the same field of endeavor of analyzing sequences of time-stamped events by machine learning in order to characterize and anticipate the behavior of the system that produced them, since (Jianwu Xu) is directed to event record logs in which "each such record can include time stamps and descriptions of the system events" (Jianwu Xu, col. 9, lines 49-51) and (Aggarwal) is directed to "a sequence of events and time of occurrence associated with each event of the sequence of events that indicates when the event occurred" (Aggarwal, ¶ [0006]), which is the same event-plus-timestamp data structure recited in Claim 1. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the event sequence analysis and clustering of (Jianwu Xu) with the joint prediction and clustering model of (Aggarwal), such that a possible future event is predicted based on a partial event sequence and the plurality of clusters of event sequences.
The motivation to combine (Jianwu Xu) and (Aggarwal) is as recited by (Aggarwal), namely that performing the prediction and the clustering jointly within a unified model yields both outputs from a common representation and thereby permits the evolution of behavior across the sequence to inform the prediction, since (Aggarwal) teaches that the unified approach "enables systems to predict both an event type and associated time of occurrence for the next event in a sequence of events and to indicate an evolution of user behavior in the sequence of events through clustering" (Aggarwal, ¶ [0005]). One of ordinary skill in the art would have been motivated to make this combination because (Jianwu Xu) is directed to fault diagnosis and expressly seeks to trace "how a fault propagates through time" (Jianwu Xu, col. 3, lines 33-36), such that anticipating the next event in a developing fault sequence, rather than merely explaining a fault after the fact, yields the predictable and desirable result of enabling intervention before the fault is complete.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HUNG VAN LE whose telephone number is (571)270-0164. The examiner can normally be reached 8 a.m. - 5 p.m..
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached at (571) 272-4128. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/HUNG VAN LE/Examiner, Art Unit 2145
/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights