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 .
Claim Rejections - 35 USC § 103
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claim(s) 1, 5-6, 8-10, 14-15, and 17-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yu et al. (From IDS: US Pub. No. 2020/0125083, published April 2020, hereinafter “Yu”) in view of Garg et al. (NPL from IDS: Temporal Attribute Prediction via Joint Modeling of Multi-Relational Structure Evolution, published March 2020, hereinafter “Garg”).
Regarding claim 1, Yu teaches a method for predicting a node state comprising:
obtaining static graphs and dynamic graphs of a plurality of nodes in a target network, wherein the static graphs and the dynamic graphs are topology views (Yu, [0041] – “As shown in FIG. 4, data preprocessor 310 processes component profiles 312 and component historical records 414 (for example, stored in a database) to determine attributed temporal graphs 416. GBPM first models the interactions between different components in a complex system as an attributed temporal graph 416. In this graph 416, each node 418 (shown individually as nodes 418-1 to 418-n) is a component, e.g. a device or a sensor, and an edge 420 represents the interaction between components.” and in [0046] – “Node-based features component 510 derives node-based features 512 from component profiles 412. These features can be categorized in two groups: static node features 514 and temporal node features 516 (for example, the features that change over time).” – teaches obtaining static and dynamic graphs of a plurality of nodes in a target network (models interactions between components in a complex system), wherein the static and dynamic graphs are topology views (data preprocessor processes profiles and record to determine attributed graphs of nodes, wherein the features of the nodes are static and dynamic));
generating spatial feature data of the plurality of nodes based on the static graphs and the dynamic graphs (Yu, [0043] – “The attributes of each node 418 are generated from the component profiles such as locations and maintenance records. This graph 416 evolves as the complex systems keep running and generating new component status records.” and in [0045] – “In the graph feature generation stage, features generator 330 processes component profiles 412 and historical records 414 using a node-based features component 510 and graph-based features component 520, respectively. These features can be grouped into two categories, (1) node-based features 512, which are derived from component profiles 412; (2) graph-based features 522, which are derived from the attributed temporal graph 416.” – teaches generating spatial feature data (location data) of the plurality of nodes based on the static graphs and the dynamic graphs (graph feature generator generates features based on graph. Features are grouped into categories, including node-based features, which is spatial feature data));
obtaining time feature data of the plurality of nodes (Yu, [0045] – “In the graph feature generation stage, features generator 330 processes component profiles 412 and historical records 414 using a node-based features component 510 and graph-based features component 520, respectively. These features can be grouped into two categories, (1) node-based features 512, which are derived from component profiles 412; (2) graph-based features 522, which are derived from the attributed temporal graph 416.” – teaches obtaining time feature data of the plurality of nodes (graph feature generator generates features based on graph. Features are grouped into categories, including temporal data, which is time feature data)),
wherein obtaining time feature data of the plurality of nodes comprises:
obtaining observation data of the plurality of nodes in a first historical time range, wherein the first historical time range is a time range before the target time range (Yu, [0036] – “Data preprocessor 310 receives raw data including component historical records with timestamps” – teaches obtaining observation data of the plurality of nodes in a first historical time range, wherein the first historical time range is a time range before the target time range);
slicing the observation data in a time dimension to obtain time slice data, wherein a data amount of the time slice data is less than a data amount of the observation data (Yu, [0051] – “Temporal graph features 526 can be derived based on periodic basis, such as per hour, day, week, month, quarter, half-year and year, etc.” – teaches slicing the observation data in a time dimension to obtain time slice data, wherein a data amount of the time slice data is less than a data amount of the observation data (features can be derived on a periodic basis, such as per hour, day, week, month, etc.)); and
obtaining the time feature data based on the time slice data (Yu, [0051] – “Temporal graph features 526 can be derived based on periodic basis, such as per hour, day, week, month, quarter, half-year and year, etc.” – teaches obtaining time feature data based on time slice data (temporal graph features derived based on per hour, day, week, month, etc.));
obtaining a predicted state of a target node in the target time range based on the spatial feature data and the time feature data, wherein the target node is any node in the plurality of nodes (Yu, [0038] – “Detector 370 is trained and tested with the features from features generator 330, such as described with respect to FIG. 6 herein below. Detector 370 can detect maintenance ready components, such as components that are appropriate to receive maintenance within a predetermined window. Components that are not maintenance ready can include those that are expected to perform above a predetermined level for the duration of the time window. The time window can be selected at a particular instance between or after a scheduled maintenance event to ensure that the component is addressed in a timely manner before issues arise.” and in [0039] – “System 300 can apply predictive maintenance that can adapt to a large volume of system running logs and component profiles and satisfy timely detection of the components that need to be replaced (or otherwise serviced).” – teaches obtaining a predicted state of a target node in the target time range (detector detects maintenance ready components for components to receive maintenance within a predetermined window) based on the spatial feature data and time feature data (detector trained and tested on feature from feature generator), wherein the target node is any node in the plurality of nodes (any component of the components detected for maintenance));
Yu fails to explicitly teach using a spatial model; using a temporal model; obtaining a true state of the target node in the target time range; and training a temporal model and a spatial model based on the true state and the predicted state, wherein the temporal model is used to output the time feature data based on input observation data, and the spatial model is used to output the spatial feature data based on the static graphs and the dynamic graphs that are input.
However, analogous to the field of the claimed invention, Garg teaches further comprising:
generating spatial feature data of the plurality of nodes based on the static graphs and the dynamic graphs using a spatial model (Garg, Section 4.1 Paragraph 1 – “We learn a head-specific representation and then model its history using Recurrent Neural Network (RNN). Let Eh,τ represent the events associated with head h at time τ, i.e., Eh,τ = {(h,ri,ti,aτ h,aτ ti ,τi)|τi = τ,∀i}. For each entity in the graph, we decompose the information in two parts, static information and dynamic information. Static in formation does not change over time and represents the inherent information for the entity. Dynamic information changes over time. It represents the information that is affected by all the external variables for the entity. For every entity h in the graph at time τ, we construct an embedding for the entity which consists of two components: 1. Static (Constant) learnable embedding (ch) which does not change over time. 2. Dynamic embedding (dh,τ = ah,τ·W1) which changes over time. where ch ∈ Rd,ah,τ ∈ Rk is the attribute of entity h at time τ and W1 ∈ Rk× d is learnable parameter… For every relation (link) r, we construct a learnable static embedding er ∈ Rd.” and in Section 4.1 Paragraph 2 – “To capture the neighbourhood information of entity h at time τ from the dynamic graph, we propose two spatial embeddings: attribute embedding (Ah,τ) and interaction embedding (Ih,τ). Ah,τ captures the spatio-attribute information from the neighbourhood of the entity h and Ih,τ captures the spatio-interaction information from the neighbourhood of the entity h. Mathematically, we can define the spatial embeddings as: Eq. (1) Eq. (2) where |Eh,τ| is the cardinality of set Eh,τ and W2 ∈ R3d× d and W3 ∈R2d× d are learnable parameters.” – teaches generating spatial feature data of the plurality of nodes based on the static graphs and the dynamic graphs using a spatial model (decomposes information into static and dynamic information, captures neighborhood information of entity h at time t by two spatial learnable embeddings . Spatial model constructs learnable spatial embeddings with learnable parameters, optimized by the graph prediction loss as in Section 4.4, thus obtaining spatial feature data using a spatial model));
obtaining time feature data of the plurality of nodes using a temporal model (Garg, Section 4.2 Paragraph 1 – “The embeddings Ih,τ, Ah,τ capture the spatial information for entity h at time τ. For predicting the information at future time, we need to capture the temporal dependence of the information. To keep track of the interactions and the attribute evolution over time, we model the history using Gated Recurrent Unit [Cho et al., 2014], an RNN. For the head s, we define the encoded attribute history at time τ as the sequence [Ah,1, Ah,2,...,Ah,τ−1] and the encoded interaction history at time τ as the sequence [Ih,1,Ih,2,...,Ih,τ−1]. These sequence provide the full information about the evolution of the head h till time τ. To represent in sequences, we model the encoded attribute and encoded interaction history for head h as follows: Eq. (3) and Eq. (4)” – teaches obtaining time feature data of the plurality of nodes using a temporal model (captures temporal dependence of information by modeling history using GRU, an RNN, thus obtaining time feature data of the nodes using a temporal model));
obtaining a true state of the target node in the target time range (Garg, Section 4.4 Paragraph 1 – “For the attribute loss, we use the mean-squared error LA = 1 N PN i=1(a 0 hi,τi −ahi,τi ) 2 , where a 0 hi,τi is the predicted attribute and ahi,τi is the ground truth attribute. For the interaction loss, we use the standard multi-class cross entropy loss, LI = PN i=1 PM c=1 yj log(P r(t = c|hi , ri)), where the M is the number classes i.e. number of entities in our case.” – teaches obtaining a true state of the target node in the target time range (ahi,τi is the ground truth of the attribute at time τ)); and
training the temporal model and the spatial model based on the true state and the predicted state (Garg, Section 4.4 Paragraph 1 – “We use multi-task learning loss for optimizing the parameters. We minimize the attribute prediction loss and graph prediction loss jointly. The total loss L =LI +λLA,whereLI is interaction loss, LA is attribute loss and λ is a hyperparameter deciding the weight of both the tasks. For the attribute loss, we use the mean-squared error LA = 1 N N i=1(ahi,τi −ahi,τi )2, where ahi,τi is the predicted attribute and ahi,τi is the ground truth attribute. For the interaction loss, we use the standard multi-class cross entropy loss, LI = N i=1 M c=1yj log(Pr(t = c|hi,ri)), where the M is the number classes i.e. number of entities in our case.” – teaches training a temporal and spatial model based on the true state and predicted state (trains temporal and spatial model on attribute prediction loss and graph prediction loss, training based on the predicted attribute and the ground truth attribute)), wherein the temporal model is used to output the time feature data based on observation data (Garg, Fig. 2 and Section 4.2 Paragraph 1 – “For predicting the information at future time, we need to capture the temporal dependence of the information. To keep track of the interactions and the attribute evolution over time, we model the history using Gated Recurrent Unit, an RNN. For the head s, we define the encoded attribute history at time τ as the sequence [Ah,1, Ah,2, . . . , Ah,τ−1] and the encoded interaction history at time τ as the sequence [Ih,1, Ih,2, . . . , Ih,τ−1]. These sequence provide the full information about the evolution of the head h till time τ . To represent in sequences, we model the encoded attribute and encoded interaction history for head h as follows: Eq. (3) and Eq. (4)” – teaches the temporal model used to output time feature data based on input observation data (encoded attribute history and encoded interaction history at time t, sequence provides full information about evolution of head h until time τ and models encoded attribute and interaction history)), and the spatial model is used to output the spatial feature data based on the static graphs and the dynamic graphs that are input (Garg, Fig. 2 and Section 4.1 Paragraph 2 – “To capture the neighbourhood information of entity h at time τ from the dynamic graph, we propose two spatial embeddings: attribute embedding (Ah,τ ) and interaction embedding (Ih,τ ). Ah,τ captures the spatio-attribute information from the neighbourhood of the entity h and Ih,τ captures the spatio-interaction information from the neighbourhood of the entity h. Mathematically, we can define the spatial embeddings as: Eq. (1) Eq. (2)” – teaches the spatial model used to output spatial feature data based on the static graphs and the dynamic graphs that are input (captures neighborhood information of entity h at time τ from the dynamic graph through two spatial embeddings, where static embeddings (ch in Eq. 2) are used to determine the interaction embedding)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the spatial and temporal models of Garg to the observation and historical time data of Yu. Doing so would capture the spatial information from knowledge graphs and provide a dynamic embedding to capture the dynamics of a time series (Garg, Introduction).
Claims 10 and 19 incorporate substantively all the limitations of claim 1 in an apparatus and non-transitory computer readable storage medium, and are rejected on similar grounds as above. Yu teaches the processor and memory of these claims at [0019] - “Each computer program may be tangibly stored in a machine-readable storage media or device (e.g., program memory or magnetic disk) readable by a general or special purpose programmable computer, for configuring and controlling operation of a computer when the storage media or device is read by the computer to perform the procedures described herein. The inventive system may also be considered to be embodied in a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.
Regarding claim 5, the combination of Yu and Garg teaches the method according to claim 1, wherein obtaining the static graphs and the dynamic graphs of the plurality of nodes in the target network comprises:
obtaining observation data of the plurality of nodes in a second historical time range, wherein the second historical time range is a time range before the target time range (Yu, [0036] – “Data preprocessor 310 receives raw data including component historical records with timestamps” – teaches obtaining observation data of the plurality of nodes in a second historical time range (multiple component historical records), wherein the second historical time range is a time range before the target time range); and
constructing the static graphs and the dynamic graphs based on the observation data (Yu, [0041] – “As shown in FIG. 4, data preprocessor 310 processes component profiles 312 and component historical records 414 (for example, stored in a database) to determine attributed temporal graphs 416.” and in “[0046] – “Node-based features component 510 derives node-based features 512 from component profiles 412. These features can be categorized in two groups: static node features 514 and temporal node features 516 (for example, the features that change over time).” – teaches constructing the static and dynamic graphs based on the observation data (processes component data to generate graphs, node features of nodes of the graphs comprise static and dynamic features)).
Claim 14 is similar to claim 5, hence similarly rejected.
Regarding claim 6, the combination of Yu and Garg teaches the method according to claim 5, wherein
the observation data comprises at least one of: meteorological data, network topology data, traffic data, voice data, signaling data, point of interest (POI) data, major-event data, or holiday data (Yu, [0035] – “System 300 implements a graph-based predictive maintenance framework, GBPM, to address the problem of detecting failure components in the domain of complex system management. Complex systems are known to consist of a great variety of components working together in a highly complex and coordinated manner and influencing each other's lifetime. For example, a cyber-physical system can be equipped with many sensors that record the status of the physical and software components. For system administrators, system 300 provides a capability to identify an optimum moment (for example, to minimize expense, maximize time, and/or properly exploit the lifetime of various components without increasing a risk of failure of that component or other related components above a predetermined threshold) when maintenance should be performed for the components so that the whole system can continue to function reliably (or to maximize output of the entire system and efficiency of usage of maintenance resources). System 300 uses GBPM to quantify the risk of failure for components in the complex systems given components' historical running records and profiles and uses this information to help schedule component maintenance. The historical running records can include mechanical properties, average usage, operating conditions, the amount of data send to or received from other components, etc.” – teaches wherein the observation data comprises at least one of network topology data).
Claim 15 is similar to claim 6, hence similarly rejected.
Regarding claim 8, the combination of Yu and Garg teaches the method according to claim 1.
Yu fails to explicitly teach the static graphs belong to first-type topology views, the dynamic graphs belong to second-type topology views, and a topological relationship change rate of the first-type topology views is less than a topological relationship change rate of the second-type topology views.
However, analogous to the field of the claimed invention, Garg teaches: wherein
the static graphs belong to first-type topology views, the dynamic graphs belong to second-type topology views, and a topological relationship change rate of the first-type topology views is less than a topological relationship change rate of the second-type topology views (Garg, Section 4 Paragraph 1 & Fig. 2 – “We model the changing graph structure and attribute values by learning an entity-specific representation. We define the representation of an entity in the graph as a combination of a static learnable embedding which does not change with time and represent static characteristics of each node, and a dynamic embedding which depends on the attribute value at that time and represent dynamically evolving property of each node” – teaches the static graphs belong to first-type topology views, and the dynamic graphs belong to second type topology views, and a topological relationship change rate of the first-type topology views is less than a topological relationship change rate of the second-type topology views (static learnable embedding which does not change with time, and dynamic embedding which represents dynamically evolving properties of each node)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the first-type and second-type topology views of Garg to the observation data and historical data of Yu. Doing so would capture the spatial information from knowledge graphs and capture the dynamics of the time series and the evolving graph (Garg, Introduction).
Claim 17 is similar to claim 8, hence similarly rejected.
Regarding claim 9, the combination of Yu and Garg teaches the method according to claim 1, wherein
the target network is a communications network (Yu, [0035] – “System 300 implements a graph-based predictive maintenance framework, GBPM, to address the problem of detecting failure components in the domain of complex system management. Complex systems are known to consist of a great variety of components working together in a highly complex and coordinated manner and influencing each other's lifetime… System 300 uses GBPM to quantify the risk of failure for components in the complex systems given components' historical running records and profiles and uses this information to help schedule component maintenance. The historical running records can include mechanical properties, average usage, operating conditions, the amount of data send to or received from other components, etc.” – teaches the target network is a communications network (complex system of components sending data to and from other components)), the static graphs are topology views representing a traffic-based topological relationship of the plurality of nodes (Yu, [0050] – “The static graph features 524 include one-hop features such as degrees, and multi-hop features. According to example embodiments, these include degree, in-degree, out-degree, in-degree/out-degree, and in-degree minus out-degree. In directed graphs, the number of edges going into a node is referred to as the in-degree of the corresponding node and the number of edges coming out of a node is referred to as the out-degree of the corresponding node. These can also include weighted version of degree features. The features can be based, for example, on 1-hop (one-hop) and 2-hop (two hop) ego net. 1-hop ego net: triangles, density, edges, number of edges entering, number of edges leaving. 2-hop ego net: triangles, density, edges, vertices, # edges enter, # edges leaving.” – teaches the static graphs are topology views representing traffic-based topological relationships of the plurality of nodes (static graph features include attributes regarding the number of edges going into and out of a node, thus representing traffic-based topological relationships of nodes)), and the dynamic graphs are topology views representing a physical-line-based topological relationship of the plurality of nodes (Yu, [0042] – “different line thicknesses are shown to illustrate the different weights associated with the intensity of interaction between nodes 418. For example, the line between node 418-0 and 418-4 is illustrated with a greater thickness (i.e., weight) than that between node 418-0 and 418-n. Note that a particular number of nodes 418, edges 420 and thicknesses are shown by way of illustration and that any number of nodes 418, edges 420 and thicknesses can be implemented.” and in [0043] – “The attributes of each node 418 are generated from the component profiles such as locations and maintenance records. This graph 416 evolves as the complex systems keep running and generating new component status records.” – teaches the dynamic graphs are topology views representing a physical-line-based topological relationship of the plurality of nodes (lines of different thickness represent intensity of interactions between nodes, attributes of each node represent component profiles and maintenance records of physical components, thus the dynamic graph is a topology view representing a physical-line-based topological relationship)).
Claim 18 is similar to claim 9, hence similarly rejected.
Claim(s) 3, 12, and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yu and Garg as applied to claims 1, 10, and 19 above, and further in view of Khoshraftar et al. (NPL from IDS: Dynamic Graph Embedding via LSTM History Tracking, published 2019, hereinafter “Khoshraftar”).
Regarding claim 3, the combination of Yu and Garg teaches the method according to claim 1.
The combination of Yu and Garg fails to explicitly teach observation data in a time range that is in the first historical time range and that is adjacent to the target time range.
However, analogous to the field of the claimed invention, Khoshraftar teaches wherein the time slice data comprises:
observation data in a time range that is in the first historical time range and that is adjacent to the target time range (Khoshraftar, Section III Subsection C, Paragraph 1 – “we define a temporal random walk as follows. Given a series of graphs G1, G2, ..., Gt a temporal neighbor walk for each node v in graph Gi is defined as a sequence of neighbors of a node v at L previous time points represented by ui−L, ui−L+1, ..., ui where ux is the neighbor of node v at time x.” – teaches observation data in a time range that is in the first historical time range and that is adjacent to the target time range (ui−L, ui−L+1, ..., ui represents a series of consecutive, adjacent historical time ranges at L previous time points)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the data in the first historical time range and adjacent to the target time range of Khoshraftar to the observation and time feature data of Yu and Garg. Doing so would reflect the changes of nodes and the structure of the graph over time (Khoshraftar, Section III Subsection C Paragraph 2).
Claims 12 and 20 are similar to claim 3, hence similarly rejected.
Claim(s) 4 and 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yu and Garg as applied to claims 1, 10, and 19 above, and further in view of Scheib et al. (US Pub. No. 2016/0359697, published Dec. 2016, hereinafter “Scheib”).
Regarding claim 4, the combination of Yu and Garg teaches the method according to claim 1.
The combination of Yu and Garg fails to explicitly teach observation data in a particular time range in the first historical time range, wherein the particular time range is located in a first period, the target time range is located in a second period, and the particular time range comprises a time range that corresponds to the target time range and that is in the first period.
However, analogous to the field of the claimed invention, Scheib teaches wherein the time slice data comprises:
observation data in a particular time range in the first historical time range, wherein the particular time range is located in a first period, the target time range is located in a second period, and the particular time range comprises a time range that corresponds to the target time range and that is in the first period (Scheib, [0047] – “the collectors 206 can retain a complete dataset describing one period (e.g., the past minute or other suitable period of time), with a smaller dataset of another period (e.g., the previous 2-10 minutes or other suitable period of time), and progressively consolidate network traffic and corresponding data of other periods of time (e.g., day, week, month, year, etc.). In some embodiments, network traffic and corresponding data for a set of flows identified as normal or routine can be winnowed at an earlier period of time while a more complete data set may be retained for a lengthier period of time for another set of flows identified as anomalous or as an attack.” – teaches observation data in a particular time range in the first historical time range (retains a complete dataset describing one period), wherein the particular time range is located in a first period (smaller dataset comprising data of a period prior to the target), the target time range is located in a second period (target time is of the first dataset), and the particular time range comprises a time range that corresponds to the target time range (smaller dataset corresponds to the first dataset) and that is in the first period (smaller dataset of a period of time prior to that of the first dataset)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the particular time ranges within periods of Scheib to the observation data and historical data of Yu and Garg. Doing so would more efficiently retain data across different periods of time rather than collecting all data indefinitely (Scheib, [0047]).
Claim 13 is similar to claim 4, hence similarly rejected.
Response to Arguments
Applicant’s arguments, see pp. 8-12 of Remarks, filed 13 February 2026, with respect to the rejection(s) of claim(s) 1-2, 5-6, 9-11, 14-15, and 18-20 under 35 U.S.C. 102(a)(1) have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made over Yu in view of Garg.
Applicant argues on pp. 8 of Remarks that Yu fails to teach “obtaining static graphs and dynamic graphs of a plurality of nodes in a target network, wherein the static graphs are dynamic graphs are topology views” and that the graphs of Yu do not segregate static relationships from dynamic ones. Examiner respectfully disagrees. The specification of the claimed invention at [0050] states “a static graph may be used as an initial graph and then may be updated by using real-time data to obtain a dynamic graph”. Thus, the dynamic graph is a static graph modified by real-time, temporal data. Yu teaches at [0041] – “As shown in FIG. 4, data preprocessor 310 processes component profiles 312 and component historical records 414 (for example, stored in a database) to determine attributed temporal graphs 416. GBPM first models the interactions between different components in a complex system as an attributed temporal graph 416. In this graph 416, each node 418 (shown individually as nodes 418-1 to 418-n) is a component, e.g. a device or a sensor, and an edge 420 represents the interaction between components.” and in [0046] – “Node-based features component 510 derives node-based features 512 from component profiles 412. These features can be categorized in two groups: static node features 514 and temporal node features 516 (for example, the features that change over time).” – teaches obtaining static graphs (attributed temporal graphs with static node features) and dynamic graphs (attributed temporal graphs with temporal node features) of a plurality of nodes (each node 418-1 to 418-n), wherein the static and dynamic graphs are topology views (interactions between components in a complex system modeled as graphs, thus a topology view of a network of components).
Applicant further argues on pp. 9 that Yu fails to teach obtaining time slice data and that the temporal aspects of Yu focus on attributing timestamps to the graphs rather than processing observation data to extract temporal patterns. Examiner respectfully disagrees. Yu teaches at [0051] – “Temporal graph features 526 can be derived based on periodic basis, such as per hour, day, week, month, quarter, half-year and year, etc.” – teaches wherein the temporal features are derived on a periodic basis, such as per hour, day, week, month, etc.. Thus, Yu teaches slicing observation data on a periodic basis, where temporal windows (per hour, day, week, etc.) are defined to slice the overall data into periodic slices, where the sliced data is a smaller data amount than the observation data. Yu further states in [0043] – “In data preprocessor 310, graphs 416 can be generated in different resolutions, e.g., hourly, daily, weekly, monthly and yearly graphs.” – which shows that Yu processes observation data (at data preprocessor 310) to extract temporal patterns.
Applicant argues on pp. 10-11 that Garg does not disclose separate training of a spatial model and temporal model. Examiner respectfully disagrees. Garg teaches at Section 4.1 Paragraph 1 – “For every entity h in the graph at time τ, we construct an embedding for the entity which consists of two components: 1. Static (Constant) learnable embedding (ch) which does not change over time. 2. Dynamic embedding (dh,τ = ah,τ·W1) which changes over time. where ch ∈ Rd,ah,τ ∈ Rk is the attribute of entity h at time τ and W1 ∈ Rk× d is learnable parameter… For every relation (link) r, we construct a learnable static embedding er ∈ Rd.” and in Section 4.1 Paragraph 2 – “To capture the neighbourhood information of entity h at time τ from the dynamic graph, we propose two spatial embeddings: attribute embedding (Ah,τ) and interaction embedding (Ih,τ). Ah,τ captures the spatio-attribute information from the neighbourhood of the entity h and Ih,τ captures the spatio-interaction information from the neighbourhood of the entity h. Mathematically, we can define the spatial embeddings as: Eq. (1) Eq. (2) where |Eh,τ| is the cardinality of set Eh,τ and W2 ∈ R3d× d and W3 ∈R2d× d are learnable parameters.” – teaches constructing learnable embeddings that capture spatial information, where the embeddings have learnable parameters, which may be optimized by the loss function of Garg in Section 4.4, and thus teaches training a spatial model. Garg at Section 4.2 Paragraph 1 – “The embeddings Ih,τ, Ah,τ capture the spatial information for entity h at time τ. For predicting the information at future time, we need to capture the temporal dependence of the information. To keep track of the interactions and the attribute evolution over time, we model the history using Gated Recurrent Unit [Cho et al., 2014], an RNN.” – capturing temporal dependence of information using a GRU, an RNN. Thus, Garg teaches training a temporal model. Lastly, in Garg at Section 4.4 – “We use multi-task learning [Ruder, 2017; Caruana, 1997] loss for optimizing the parameters. We minimize the attribute prediction loss and graph prediction loss jointly” – teaches using multitask learning to minimize a joint loss, thus there are distinct loss functions for the spatial and temporal models that minimized jointly. The claim, as drafted, requires “training the temporal and spatial model…”, and thus does not exclude minimizing a joint loss for training the models.
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., segregated static relationships (e.g. based on physical connection) from dynamic ones (e.g. based on time-varying interactions); training separate spatial model and temporal model) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Teimury et al. (NPL: Estimation of time-series on graphs using Bayesian graph convolutional neural networks, published Sept. 2019) teaches reconstruction of time-series on graphs to determine static and dynamic features. Teaches using a graph convolutional neural network to obtain spatial features. Teaches using a gated recurrent unit to obtain temporal features. Teaches using a combination of GCN and GRU to learn the temporal correlation from dynamic and spatio-temporal datasets. Uses GCN and GRU to extract spatial and temporal features, respectively.
Zhao et al. (NPL: T-CGN: A Temporal Graph Convolutional Network for Traffic Prediction, published Dec. 2018) teaches accurate, real-time traffic forecasting using a neural network-based traffic forecasting method. Teaches a temporal graph convolutional network model, which is a combination of the graph convolutional network (GCN) with a gated recurrent unit (GRU). Teaches using the GCN to capture spatial dependence and using the GRU to capture the dynamic change of traffic flow to model the temporal dependence.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
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, MATT ELL can be reached at 571-270-3264. 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.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141
/TAN H TRAN/Primary Examiner, Art Unit 2141
Prosecution Insights