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 .
This action is in response to the amendment filed 2/20/26. Claims 1-20 are pending in this action. Claims 1-20 have been amended.
Claim Objections
Claims 1-14 are objected to because the phrase, “a method” is repeated in the preamble. Correction is requested.
Claim Interpretation
Certain claim limitations are interpreted in view of the following sections of applicants’ disclosure.
Low signal datasets- described in
Applicants’ disclosure [0002] In view of the aforementioned problems, novel methods and systems are described herein for generating a trained machine learning model from low signal datasets. Many machine learning models are generated by inputting a dataset into a training routine such that the training routine generates parameters for the features in the training dataset. As discussed above, this may not work well for datasets that have small sample sizes or have class imbalance, sometimes referred to as low signal datasets. The disclosed approach uses Bayesian Hierarchical model that is used to generate posterior distribution for parameters of the model based on priors and training dataset. Training dataset is enhanced by including a plurality of entries from similar segments as the segment with low-signal dataset. For example, while modeling probability of default (PD) for healthcare companies, data from generic commercial companies is included in the training dataset.
Segment – described in
Applicants discourse [0017] ML training system 102 may receive input from an expert to generate a probability distribution for a parameter. The probability distribution may include a plurality of values and a plurality of probabilities. FIG. 2 illustrates one possible probability distribution for a given parameter. Columns 203, 206, and 209 of data structure 200 may illustrate different values for a particular parameter and associated probabilities for each value. In one example, the feature may be liquidity and the mechanism described in this disclosure may be used to generate a Bayesian Hierarchical model that accurately predicts default in low-default portfolios. Each portfolio may represent a segment (e.g., a type of company) such that each segment has industry-specific risk drivers that affect default. These segments may include healthcare providers, commercial and industrial companies, energy companies, etc. Thus, to get accurate default predictions, parameter sensitivity is required for each segment (e.g., each type of company) so that the specifics of each industry are captured. In some embodiments, the inputs for prior distribution or prior distributions may be received from an expert or multiple experts. Those inputs may be stored in a database and retrieved by ML training system 102.
Append – described in
Applicants’ disclosure [0025] In some embodiments, ML training system 102 may add segment data and/or macroeconomic data to the training dataset. For example, for a machine learning model being trained for healthcare companies, training system may add data relating to the healthcare industry. Macroeconomic data may include disposable income per capita, personal consumption of services, number of bankruptcies, energy prices (e.g., crude oil prices) and/or other factors. In some embodiments, segment and/or macroeconomic data may be reported at a different time interval than the training data. ML training system 102 may append the data taking care of the time difference to create the updated training dataset. For example, segment and/or macroeconomic data may be collected quarterly while training dataset may be collected yearly. The data in the training dataset and segment/macroeconomic data have a timestamp associated with a collection time of the data. Accordingly, ML training system 102 may determine, for each entry in the training dataset, a corresponding time period of the plurality of time periods to add segment/macroeconomic data to each entry within the updated dataset.
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 1-5, 8-12, and 15-19 are rejected under 35 U.S.C. 103 as being unpatentable over Sun et al. (US 2019/0080246 A1; hereinafter "Sun") in view of US PG-PUB 2021/0073669 issued to Banerjee et al., hereinafter “Banerjee,” further in view of Yuan et al. (US 2020/0019984 A1; hereinafter "Yuan")
As per Claim 1, Sun teaches a method for generating a trained machine learning model
using low signal datasets, the method comprising:
receiving a first training dataset comprising a first plurality of features and a first
plurality of entries, wherein the first training dataset comprises first data for a plurality of
segments to be modeled (Sun [0070]: "Referring now to FIG. 2, the Bayesian modeler 122 is
shown in greater detail according to an illustrative implementation. The brand data 130 is
shown to be an input to the category model generator 124. Based on the brand data 130 and
the weak priors 132, the category model generator 124 can be configured to generate the HCM
134. The brand data 130 is shown to be brand data for three separate brands, i.e., brand data
set 1, brand data set 2, and brand data set 3, however, there may be any number of brands.
The Brand data set 1 is shown to include media channels 202, content inputs 204, responses
206, and time 208. The brand data set 1 may include data for one or more media channels 202
e.g., Internet advertising media channel data, radio media channel data, printed advertising,
etc. The content inputs 204 may indicate a particular content input for one of the media
channels 202 at a particular time, e.g., a particular amount of advertising spend, a particular
number of advertisement impressions, etc. The responses 206 may indicate a total response for
the brand data 1 at a particular time 208. For example, the response 206 may be total revenue
for the brand, number of online conversions for the brand, number of registrations for a
subscription, etc."
Here, is received a first training dataset ("brand data set 1") comprising a first plurality
of features ("media channels", "content inputs", "responses", "time") with a first plurality of
entries (each element in the data set) and a first data ("brand data set 1") for a plurality of
segments to be modeled ("brand data set 1, brand data set 2, and brand data set 3".))
receiving a second training dataset comprising a second plurality of entries, wherein
the second plurality of entries comprises second data for a segment of the plurality of
segments to be modeled (Sun [0070] above, for "brand data set 2"). Examiner note: A second
training dataset is also taught by Yuan below.
receiving feature groups for selected features and inputs to generate a prior
probability distribution for parameters of the feature groups for the selected features,
wherein the prior probability distribution comprises a plurality of values and a plurality of
probabilities (Sun [0091]: "Referring now to FIG. 5B is a set of two charts, chart 550 and 552,
that illustrate similarity across brands in terms of price and sales according to an illustrative
implementation. Impact of price is shown in chart 550 while media sales for certain brands is
shown in chart 552. This may support using a shared prior distribution on media parameters
across similar brands."
Here, a feature group (features of "similar brands") for selected features ("price and
sales") is used to generate a prior probability distribution ("shared prior distribution") for
parameters ("on media parameters") of the feature groups (the parameters for the brands
which have the price and sales features by which they are grouped, are therefore parameters
"of" the feature groups.
Examiner further notes that a plurality of values with a plurality of associated
probabilities is merely the definition of a probability distribution. This is shown on the drawings
of priors on Sun Fig. 4.).
[As for the “segment” recited in the step of appending, Sun teaches, Sun teaches category, brand, competitor (i.e., company related information), each of them reads on claimed “segment.” See also Sun, [0057] One further advantage of category analysis is the ability to incorporate competitive factors, e.g., impact across brands, into an MMM (Media Mix Model). Developing MMMs for a single brand can suffer from omitted variables, of which competitive factors, such as competitor price and promotion, are common ones. In category analysis, impact from competitor activities on a brand of interest can be explicitly included in the model to help reduce bias in parameter estimates.]
Sun also teaches a Bayesian hierarchical model, in [0074] Referring now to FIG. 3, a process 300 for generating the HCM 134 and the HBM 138 is shown according to an illustrative implementation. The analysis system 120 can be configured to perform process 300. Specifically, the components of the analysis system 120, i.e., Bayesian modeler 122, the category model generator 124, and the brand model generator 126 can be configured to perform process 300. Further, any computing device described herein can be configured to perform process 300, e.g., the computing system 3600 of FIG. 36.
[0130] FIG. 8 shows that informative priors derived from the category model can help improve the estimation accuracy and reduce uncertainty by passing on learnings obtained from a richer dataset. The compromise between data and priors is a standard example of the posterior distribution as a result of both the data and the priors: when the information in the data is weak, an informative prior has more influence on the posterior. Using the category dataset to derive informative priors can supplement the lack of information content of a single brand dataset.]
With respect to claim 1, Sun does not explicitly indicate the step of “appending, based on a training routine of a machine learning model for a segment of the plurality of segments and based on the first data including low signal datasets for the segment of the plurality of segments, the first training dataset with a portion of the second training dataset to generate an updated training dataset;
Banerjee, in a method for generating training data for machine-learning models uses an augmented dataset. Banerjee, in Fig 2, shows an original dataset (216), an augmented and/or updated dataset (219) with new records (229).
Banerjee recognizes the problem of accuracy prediction and bias caused by the small datasets (see Banerjee, [0002] sufficiently large data sets are not always readily available for use in training a machine learning model. For example, tracking an occurrence of an event that occurs rarely may lead to a small dataset due to a lack of occurrences of the event”) and suggest to generate additional records to expand the small dataset (see Banerjee, (0011] However, according to various embodiments of the present disclosure, it is possible to generate additional
records that are sufficiently indistinguishable from previously collected data present in the small dataset. As a result, the small dataset can be expanded using the generated records to a size sufficient to train a desired machine learning model (e.g., a neural network, Bayesian network,
sparse machine vector, decision tree, etc.)”
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 teachings of Banerjee with Sun to improve the operation of Sun’s Bayesian Hierarchical Model (see Abstract, Sun) to make the combined system more efficient with safeguard against bias and improved accuracy.
However, Sun and Banerjee does not explicitly teach arranging the parameters
into groups based on the feature groups and generating a common prior probability
distribution for the parameters in one or more groups of the feature groups; and training the
machine learning model using the updated training dataset, wherein training the machine
learning model comprises updating the common prior probability distribution for the
parameters.
Yuan teaches wherein a training routine of a machine learning model updates the first
training dataset with a portion of the second data from the second training dataset to
generate an updated training dataset (Yuan [0051-0052]: "The model training module 304 is
representative of functionality to fit a statistical model to the data 308 through use of the
Dynamic Hierarchical Empirical Bayes (DHEB) techniques described herein. The prediction
generation module 306 is representative of functionality to generate a prediction based on the
statistical model and corresponding hierarchical structure generated by the model training
module 304 to process subsequently obtained data. In an implementation, the model training
module 304 is configured to refresh the hierarchical structure of the statistical model at
predefined intervals because in practice the hierarchical structure does not exhibit significant
amounts of changes over relative short periods, e.g., daily. Thus, instead of refreshing the
structure every day, the system is designed to retrain the hierarchical structure of the statistic
model based on a parameter T which define a period of hierarchy updating. For example, T=1
may cause the hierarchical structure to be updated daily, T=2 specifies an update every other
day, and so on. Between updates, changes to the statistical model are "frozen" and thus newly
collected data 308 flows through the hierarchy and used to obtain prediction at each node at a
decision unit level. In practice, it has been found that a value of T as equal to four reduces time
complexity and supports real time output without sacrificing accuracy."
Here, Yuan discloses receiving a second training data set ("subsequently obtained
data"), and adding this to the first set of training data to form an updated training set ("model
training module 304 is configured to refresh the hierarchical structure of the statistical model at predefined intervals") arranging the parameters into groups based on the feature groups and generating a common prior probability distribution for the parameters in one or more groups of the feature groups (Yuan [0076]: "In this section, a Dynamic Hierarchical Empirical Bayes technique is used to generate a hierarchical structure using a hierarchical shrinkage loss as described in the previous section. In a multi-level hierarchical Bayesian technique, parameters of child nodes under a same parent node are taken from a common prior distribution and the prior information flows through the hierarchical structure. In a full Bayesian analysis, a complete joint posterior distribution is generated according to the pre-determined hierarchy, and simulations are usually applied to generate inferences." Yuan [0078]: "In the illustrated
example, assume "Geo" is used as the splitting variable to split data for each "Geo" as child
nodes 616, 618, 620." Yuan [0079]: "An assumption is made for each of the ß.sub.js across
different "Geo"s are related and generated from a common prior distribution because these
nodes share the same parent nodes, which isB.sub.j.sub.priorN{u.sub.0,g.sub.0.sup.2] Based
on the previous prior discussion, a posterior distribution of ß.sub.j is obtained for each "Geo"
node 616, 618, 620 as follows."
Here, Yuan discloses arranging the parameters into groups ("child nodes under a same
parent node") based on the feature groups (based on grouping by the feature "Geo"), and then
generating a common prior probability distribution ("common prior distribution")).
Examiner also notes that not only parent and child nodes may be grouped with a
common prior distribution, but also those that are simply "nearby" as stated in Yuan [0066]: "In
one example, a multi-level hierarchical Bayesian technique is used to propagate information
across the hierarchical structure and allow information to be shared among sub groups of
nodes that are nearby to each other in the hierarchical structure. Decision units, for instance, in the same ad groups may perform in a similar manner and thus may share a same prior
distribution.") training the machine learning model using the updated training dataset, wherein
training the machine learning model comprises updating the common prior probability
distribution for the parameters (Yuan [0002]: "One technique used to address this challenge in
data sparsity leverages the rest of the information in the big data, which is referred to as
Hierarchical Bayes (HB). In Hierarchical Bayes, a statistical model includes multiple levels that
form a hierarchical structure that is used to estimate parameters of a posterior distribution
using a Bayesian method. Bayes theorem is used to integrate sub-models to form the
hierarchical structure and account for uncertainty. The hierarchical structure of the statistical
model is then used to update a probability estimate as additional evidence on a prior
distribution is received in order to form a prediction based on past observances." Yuan [0051-
0052]: "The model training module 304 is representative of functionality to fit a statistical
model to the data 308 through use of the Dynamic Hierarchical Empirical Bayes (DHEB)
techniques described herein. The prediction generation module 306 is representative of
functionality to generate a prediction based on the statistical model and corresponding
hierarchical structure generated by the model training module 304 to process subsequently
obtained data. In an implementation, the model training module 304 is configured to refresh
the hierarchical structure of the statistical model at predefined intervals."
Here, Yuan discloses that the hierarchical structure is used to "update a probability
estimate as additional evidence on a prior distribution", and goes on to describe that the
training is to "fit a statistical model" and "refresh" the structure of the statistical model at
"intervals" based on the updated training dataset ("subsequently obtained data".))
Yuan is analogous art because it is in the field endeavor of Hierarchical Bayes, as is also
Sun (Abstract: "Systems, methods, and computer-readable storage media that may be used to
generate a category Bayesian hierarchical model"). It would have been obvious before the
effective filing date of the claimed invention to combine the Hierarchical Bayes approach of Sun with the common prior distribution of grouped parameters of Yuan. One of ordinary skill in the art would have bene motivated to do so in order to address the problem of data sparsity, which is acknowledged by Sun (Sun [0049]: "One approach to address the problem of data sparsity is to inject variability through randomized experiments") and is directly addressed by Yuan (Yuan [0065]: "Application of the same prior distribution for all ß.sub.is and use of a posterior mean as the predicted RPC for each bid unit results in non-sparse predictions due to borrowing of information by incorporating a prior distribution.")
As per Claim 2, the combination of Sun and Banerjee and Yuan teaches the method of claim 1. Yuan
teaches further comprising generating the prior probability distribution that represents
domain expertise prior to updating the common prior probability distribution for the
parameters by training the machine learning model. (Yuan [0051-0052]: "The model training
module 304 is representative of functionality to fit a statistical model to the data 308 through
use of the Dynamic Hierarchical Empirical Bayes (DHEB) techniques described herein. The
prediction generation module 306 is representative of functionality to generate a prediction
based on the statistical model and corresponding hierarchical structure generated by the model
training module 304 to process subsequently obtained data. In an implementation, the model
training module 304 is configured to refresh the hierarchical structure of the statistical model at
predefined intervals."
Examiner notes that the Specification provides no detail on what constitutes "domain
expertise". The broadest reasonable interpretation of this limitation is simply that the first
training iteration of Sun's iterative process amounts to "domain expertise".)
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 teachings of Yuan with Sun and Banerjee for at least the reasons recited in the rejection to Claim 1.
As per Claim 3, the combination of Sun and Banerjee and Yuan teaches the method of claim 1. Sun
teaches further comprising: receiving segment data comprising data entries for a plurality of
time periods; determining, for each entry in the updated training dataset, a corresponding
time period of the plurality of time periods; and adding, to each entry in the updated training
dataset, a corresponding portion of the segment data associated with each corresponding
time period of the plurality of time periods. (Sun [0068]: "The brand data 130 may include sets
for different brands that each include responses, content inputs, a content types, and/or time
identifiers. The content type may indicate a particular media channel of the set of data, for
example, television, radio, Internet advertising, a particular advertising campaign, etc. The
response may indicate particular amounts of revenue at particular times. In some
embodiments, the response is number of conversions, number of sales, etc. The content inputs
may indicate particular amounts of advertising spending for the content type at particular
times. The content inputs may indicate a number of advertisements run. The time identifiers
may indicate that there was a particular amount of response and content input for a particular
content type for a particular brand. In this regard, the brand data 130 may be stored as time
based vectors."
Examiner notes that here, all of the received data for a segment may comprise a time
identifier, which means that entries may be for a plurality of time periods, and each entry is for
a particular time period, and therefore corresponding portions of the data are for
corresponding time periods.)
As per Claim 4, the combination of Sun and Banerjee and Yuan teaches the method of claim 1. Sun
teaches wherein generating the machine learning model comprises generating a function for
one or more parameters, wherein the function comprises the plurality of values, and wherein
each value of the plurality of values is associated with a probability. (Examiner notes that a
function that maps a plurality of values each to a respective probability is merely the definition
of a probability distribution. Sun Figure 4 discloses graphs of functions that represent
probability distributions. Sun [0091] discloses that the prior distributions are for parameters:
"This may support using a shared prior distribution on media parameters across similar
brands.")
As per Claim 5, the combination of Sun and Banerjee and Yuan teaches the method of claim 1. Sun
teaches wherein the training routine of the machine learning model uses a posterior
distribution to generate probability distribution of model output for a plurality of entries.
(Sun [0081]: "The generated HBM 138 may include one or more fitting model parameters (e.g.,
generating posteriors for the model parameters). Based on the fitted model parameter (e.g.,
posteriors), the brand model generator 126 can be configured to generate the ROAS and the
mROAS.")
Claim 8 is a system claim corresponding to method Claim 1, which also recites one or
more processors and a non-transitory computer-readable storage medium. Sun [0229]
teaches: "According to various implementations, the processes that effectuate illustrative
implementations that are described herein can be achieved by the computing system 3600 in
response to the processor 3610 executing an arrangement of instructions contained in main
memory 3615. Such instructions can be read into main memory 3615 from another computer readable
medium, such as the storage device 3625." Therefore, Claim 8 is rejected for similar
reasons as Claim 1.
Claim 9 is a system claim corresponding to Claim 2, and is rejected for similar reasons.
Claim 10 is a system claim corresponding to Claim 3, and is rejected for similar reasons.
Claim 11 is a system claim corresponding to Claim 4, and is rejected for similar reasons.
Claim 12 is a system claim corresponding to Claim 5, and is rejected for similar reasons.
Claim 15 is a non-transitory computer readable storage medium claim corresponding to
method Claim 1, which also recites one or more processors and a non-transitory computer readable
storage medium. Sun [0229] teaches: "According to various implementations, the
processes that effectuate illustrative implementations that are described herein can be
achieved by the computing system 3600 in response to the processor 3610 executing an
arrangement of instructions contained in main memory 3615. Such instructions can be read
into main memory 3615 from another computer-readable medium, such as the storage device
3625." Therefore, Claim 15 is rejected for similar reasons as Claim 1.
Claim 16 is a non-transitory computer readable storage medium claim corresponding to
Claim 2, and is rejected for similar reasons.
Claim 17 is a non-transitory computer readable storage medium claim corresponding to
Claim 3, and is rejected for similar reasons.
Claim 18 is a non-transitory computer readable storage medium claim corresponding to
Claim 4, and is rejected for similar reasons.
Claim 19 is a non-transitory computer readable storage medium claim corresponding to
Claim 5, and is rejected for similar reasons.
7. Claims 6, 13, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over the
combination of Sun and Banerjee and Yuan, further in view of Bai et al. ("The Inverse Gamma-Gamma Prior
for Optimal Posterior Contraction and Multiple Hypothesis Testing"; hereinafter "Bai").
As per Claim 6, the combination of Sun and Banerjee and Yuan teaches the method of claim 1 as well
as wherein arranging the parameters into the groups based on the feature groups and
common prior probability distribution (see Yuan in the rejection to Claim 1). Sun teaches
wherein arranging the parameters into the groups based on the feature groups comprises:
standardizing continuous variables (Sun [0065]: "For example, brand A may have
various content inputs (e.g., advertising spending amounts) and responses (e.g., revenue) for a
one or more of points in time (e.g., the data may be a time series) for one or more media
channels (e.g., Internet, television, radio, printed publications)." Sun [0076]: "In some
embodiments, the Bayesian modeler 122 can be configured to scale the brand data 130. For
example, the Bayesian modeler 122 can be configured to scale (e.g., normalize) the responses
(e.g. response 206) and the content inputs (e.g., content inputs 204) of the brand data 130. The
Bayesian modeler 122 can be configured to scale the content input of each brand data set
between zero and one. The Bayesian modeler 122 can be configured to use Equation 3
described herein to perform normalization. This scaled brand data 130 can be used by the
category model generator 124 to generate the HCM 134. Likewise, the scaled brand data 130
(e.g., a scaled version of a particular brand data set) can be used by the brand model generator
126 to generate the HBM 138."
Here, Sun discloses a continuous variable ("responses (e.g., revenue)") and discloses
standardizing that variable ("scale (e.g., normalize) the responses")).
and modeling prior distributions of the continuous variables using normal distribution
centered around zero (Sun [0100-0101]: "To simplify the specification of weak priors (e.g., the
weak priors 132) for custom-character across different media, media variables (e.g., the
content inputs 204 and/or responses 206 for the media channels 202) can first be scaled to be
between 0 and 1 Then, a first MMM that allows for geometric carryover effects and a flexible
shape structure, can be written as and Et
~ Normal
(0, s²))."
Here, Sun, in the process or specifying a prior distribution, as some point makes use of a
normal distribution centered around zero ("Normal(0, s²)").
However, Sun and Banerjee do not teach wherein the modeling comprises modeling a common
group variance using inverse gamma-gamma distribution.
Bai teaches modeling prior distributions of the continuous variables using normal
distribution centered around zero, wherein the modeling comprises modeling a common
group variance using inverse gamma-gamma distribution (Recall above that Yuan discloses a
common group prior probability distribution in the rejection to Claim 1. Bai, Page 1 Abstract: "In
this article, we introduce a new fully Bayesian scale-mixture prior known as the inverse gammagamma
(IGG) prior." Bai, Page 6 Section 2 "The Inverse Gamma-Gamma (IGG) Prior", discloses:
"Suppose we have observed X ~ N(Q, In), and our task is to estimate the n-dimensional vector, q.
Consider putting a scale-mixture prior on each Qn, i = 1, n of the form
(5)
We reparametrize (5) as follows:
)
(8)
Examiner notes that here, Bai discloses using a normal distribution centered around zero (N(0,
Xisi)) and modeling a variance (X|}) using inverse gamma-gamma distribution.)
Bai is analogous art because it is in the field of endeavor of Bayesian inference. It would
have been obvious before the effective filing date of the claimed invention to combine the
hierarchical bayes model of Sun and Banerjee and Yuan with the inverse-gamma-gammal prior of Bai. One of
ordinary skill in the art would have been motivated to do so in order to address the data
sparsity problem in a way that improves performance (Bayes, Abstract: "We study the well-known
problem of estimating a sparse n-dimensional unknown mean vector = (1, n) with
entries corrupted by Gaussian white noise We illustrate through simulations and data
analysis that the IGG has excellent finite sample performance for both estimation and
classification.")
Claim 13 is a system claim corresponding to Claim 6, and is rejected for similar reasons.
Claim 20 is a non-transitory computer readable storage medium claim corresponding to
Claim 6, and is rejected for similar reasons.
8. Claims 7 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over the
combination of Sun and Banerjee and Yuan, further in view of Zhang et al. ("A Hierarchical Bayes-Based
Evolutionary Ensemble Classification Algorithm"; hereinafter "Zhang").
As per Claim 7, the combination of Sun and Banerjee and Yuan teaches the method of claim 1.
However, the combination does not teach further comprising generating, based on a posterior
distribution, a plurality of classifications as output for each entry by mapping a posterior
predictive distribution to a range of classes, wherein each class in the range of classes is
associated with a value representative of a corresponding range.
Zhang teaches further comprising generating, based on a posterior distribution, a
plurality of classifications as output for each entry by mapping a posterior predictive
distribution to a range of classes, wherein each class in the range of classes is associated with
a value representative of a corresponding range. (Zhang, Abstract: "This paper proposed a
Hierarchical Bayes-based Evolutionary Ensemble (HBEE) classification algorithm that computes
and utilises our new data-driven posterior-based class similarity to evolve a tree of weak
classifiers." Zhang, Page 140: "Therefore, the original multi-label classification problem has
been transformed into a series of parallel binary classification problem Given the sufficient
ensemble basis generated in subsection III-A, category difference information can be encoded
into a posterior-based matrix (P-matrix), under the basis of the ensemble pool, which
transforms the data from original data feature space into the space spanned by the weak SVM
basis."
Examiner notes that here, Zhang teaches producing a plurality of classifications as
output ("series of parallel binary classification") by mapping a posterior distribution to a range
of classes ("posterior-based matrix under the basis of the ensemble pool"), wherein each
class in the range of classes is associated with a value representative of a corresponding range
(each is a binary classification, and thus falls into a range on one or another side of a decision
boundary.))
Zhang is analogous art because it is in the field of endeavor of Hierarchical Bayes and,
like Sun and Banerjee and Yuan, addresses a scarcity of data (Zhang, Abstract: "imbalanced data"). It would
have been obvious before the effective filing date of the claimed invention to combine the
hierarchical bayes models of Sun and Banerjee and Yuan with the hierarchical bayes-based classifier of Zhang.
One of ordinary skill in the art would have been motivated to apply hierarchical bayes to a
classification problem because it is insensitive to imbalanced class distribution of training data
(Zhang, Abstract: "The posterior information is gathered from a reduced Bayes theorem, which
is insensitive to imbalanced data amount and imbalanced inter-class similarities.")
Claim 14 is a system claim corresponding to Claim 7, and is rejected for similar reasons.
Prior Art of Record
9. The prior art made of record and not relied upon is considered pertinent to applicant's
disclosure.
Response to Arguments
Applicant’s arguments with respect to Claims 1-5, 8-12, and 15-19 being rejected under 35 U.S.C. § 103 as being unpatentable over SUN (US Patent Application Publication No. 2019/0080246) and YUAN (US Patent Application Publication No. 2020/0019984), and claims 6, 13, and 20 being rejected under 35 U.S.C. § 103 as allegedly being unpatentable over SUN, YUAN, and BAI ("The Inverse Gamma-Gamma Prior for Optimal Posterior Contraction and Multiple Hypothesis Testing), and claims 7 and 14 are rejected under 35 U.S.C. § 103 as allegedly being unpatentable over SUN, YUAN, and ZHANG ("A Hierarchical Bayes-Based Evolutionary Ensemble Classification Algorithm") have been considered but are moot because the new ground of rejection adds a new reference, Banerjee, to reject the amended claims.
Applicant amended the claims by adding, “appending, based on a training routine of a machine learning model for a segment of the plurality of segments and based on the first data including low signal datasets for the segment of the plurality of segments, the first training dataset with a portion of the second training dataset to generate an updated training dataset;” in all independent claims.
The Sun reference teaches the “segment” and the Banerjee reference teaches the “appending” step. The teachings have been detailed in the 103 rejection set forth in this action.
Applicant argues that “the applied references do not disclose or suggest "appending, based on a training routine of a machine learning model for a segment of the plurality of segments and based on the first data including low signal datasets for the segment of the plurality of segments, the first training dataset with a portion of the second training dataset to generate an updated training dataset," as recited in amended claim 1.
Please refer to the teachings of the Banerjee reference.
Applicant also argues, the Sun reference, “does not disclose "wherein a training routine of a machine learning model updates the first training dataset with a portion of the second data from the second training dataset to generate an updated training dataset." Applicant also states that the Office Action relies on paragraphs 51 and 52 of YUAN.
Please refer to the teachings of the Banerjee reference.
Applicants’ argument about the Yuan reference “merely, "refresh[ing] the hierarchical structure of the statistical model at predefined intervals" does not disclose specifically "appending, based on a training routine of a machine learning model for a segment of the plurality of segments and based on the first data including low signal datasets for the-11-
segment of the plurality of segments, the first training dataset with a portion of the second training dataset to generate an updated training dataset," as recited in amended claim 1 (emphasis added)” is moot, because the limitation is being addressed with the teachings of the Banerjee reference.
Conclusion
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.
Contact Information
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HOSAIN T ALAM whose telephone number is (571)272-3978. The examiner can normally be reached Mon-Thu, 8:00 - 4:30.
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.
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.
/HOSAIN T ALAM/Supervisory Patent Examiner, Art Unit 2132
Prosecution Insights