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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on April 30th, 2025 has been entered.
Response to Arguments
Applicant’s other arguments with respect to claims have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Examiner’s Remarks
Generative Adversarial Networks (GANs) are a type of deep learning system that uses two competing neural networks to create new, realistic data that resembles a training set. The generator creates new data (like images), while the discriminator tries to tell the difference between real data and the fake data from the generator. This competitive process allows the generator to produce increasingly convincing outputs over time. Fox uses a GAN as part of its network and inherently teaches as part of its process “generating one or more counterfactual data sets for the test data set,” “generating… a predefined number of counterfactual data sets;” “training the abnormality detection model based on the feature weights and training data sets; removing abnormal data from the training data sets; “and training, after the removing the abnormal data from the training data sets, the abnormality detection model to be a counterfactual generation model that is configured to generate the one or more counterfactual data sets.”
In particular, Goodfellow teaches that the core operation of a GAN is to repeatedly generate synthetic examples from a latent distribution and compare them against real training data so that the generator learns to approximate the underlying data manifold while the discriminator learns to distinguish realistic from unrealistic samples. In doing so, a GAN naturally produces what the present claims call “counterfactual data sets,” i.e., alternative versions of the input data that do not actually occur in the training set but are sampled from the learned distribution. Because training proceeds by back-propagating an error signal through the discriminator and into the generator, the model is, by definition, “trained based on feature-weights and training data sets” in the sense recited, where the feature weights are the learned parameters of the generator/discriminator networks and the training data sets are the real examples used to drive the adversarial learning process.
Moreover, as explained in Di Mattia, a standard use case of GANs is data cleaning and anomaly detection, where samples that the discriminator scores as abnormal or low-likelihood relative to the learned distribution are identified as outliers and removed or down-weighted from subsequent training rounds. It is therefore routine to “remove abnormal data from the training data sets” and then continue training the same underlying network so that it better models the distribution of normal data only. Once trained in this way, the generator network functions as a “counterfactual generation model” because, given a particular input condition (e.g., a nominal or partially abnormal test sample), it generates one or more synthetic alternative outputs that represent plausible non-abnormal variants of that sample. Thus, the claimed steps of generating multiple counterfactual data sets, training using feature-weights and training data, removing abnormal data, and retraining a model to generate counterfactuals are simply the routine, expected activities of a GAN-based abnormality detection system as taught by Goodfellow and implemented in Fox.
See attached NPLs: Generative Adversarial Networks, Goodfellow et al, Goodfellow, 2017; A Survey on GANs for Anomaly Detection, Di Mattia, arXiv, 2019.
The PC (Peter-Clark) algorithm is a cornerstone constraint-based method for causal discovery, starting with a fully connected undirected graph and pruning edges using statistical independence tests to find a skeleton (undirected graph) and then orienting edges to form a CPDAG (partially directed acyclic graph). Its pairwise steps involve removing edges between marginally independent variables (empty set tests) and then conditionally independent pairs (conditioning on other variables). These tests build the skeleton, after which unconnected edges are given directions help orient edges to find causal paths, aiming to uncover the underlying causal structure from observational data.
The PC algorithm is used in Liao and inherently teaches: “determining a causal relationship of the abnormality based on the quantitative feature dependence via a causality discovery algorithm;” “and generating a causal graph that represents the causal relationship of the abnormality, wherein the causality discovery algorithm is a pairwise causality discovery algorithm, and the determining of the causal relationship of the abnormality based on the quantitative feature dependence comprises;” and claim 2 “implementing an action to mitigate the abnormality based on the causal graph.”
Because of the way the PC algorithm is defined, the steps recited in the claim are simply the ordinary operations of a PC-style causal discovery routine. PC takes as input quantitative observations of multiple features (including any “abnormality” variable) and, for every pair of variables, applies statistical (in)dependence tests such as correlation, partial correlation, χ², or similar test statistics. Those tests are explicitly measures of “quantitative feature dependence,” and the outcome of each test (rejecting or failing to reject independence) determines whether the edge between the corresponding pair of variables is retained or removed in the evolving graph. As these tests are repeated while conditioning on different subsets of the remaining variables, the algorithm is, by construction, a “pairwise causality discovery algorithm” that evaluates, for each candidate pair, whether there is sufficient quantitative dependence to support a causal connection under the constraint-based assumptions.
Once the skeleton has been obtained and the orientation rules are applied, the PC algorithm outputs a CPDAG in which the remaining edges and their directions encode the discovered causal relationships between the variables. When one of those variables represents an abnormality and the others represent quantitative features, the resulting CPDAG is, by definition, “a causal graph that represents the causal relationship of the abnormality,” because the abnormality node is connected only to those feature nodes for which the quantitative dependence tests indicated a non-independent, causally compatible relationship. Thus, using PC in Liao to test quantitative feature dependence, prune and orient edges, and output a CPDAG is nothing more than performing the conventional causal-graph generation that the claims describe in terms of determining a causal relationship of the abnormality via a causality discovery algorithm and generating a causal graph that represents that relationship.
See attached NPL: Review of Causal Discovery Methods Based on Graphical Models, Frontiers in Genetics, Glymour et al; Glymour.
Allowable Subject Matter
Claims 8-13 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
The prior art of record, whether considered individually or in any combination, fails to teach or suggest the particular end-to-end, quantitative “feature dependence” determination and reporting pipeline recited in claims 8-9 and further fails to teach or suggest the specific training regimen for the feature-dependence interpretation model recited in claims 10–13.
The applied references do not teach or suggest: (i) determining a quantitative feature dependence between a test data set and one or more counterfactual data sets by (ii) calculating, for each feature of abnormal data, a numerical difference between the abnormal feature value and a median value of the counterfactual data set(s); (iii) generating feature contributions for that numerical difference “via a feature dependence interpretation model”; (iv) eliminating feature-dependence interpretations whose contributions are below a threshold; and (v) transforming the remaining feature contributions into a table that quantitatively describes feature dependence. The art may describe attribution/importance generally, and may describe baselines or counterfactual concepts generally, but it does not disclose this particular median-based counterfactual differencing coupled to a feature-dependence interpretation model and subsequent threshold-based pruning to yield a quantitative feature-dependence table in a pipeline that requires a GAN and a causal discovery algorithm. Such a pipeline is not obvious to one of ordinary skill in the art.
The pipeline fuses (i) counterfactual-baseline construction (median of counterfactual data sets), (ii) explanation/attribution modeling to generate “feature contributions,” (iii) threshold-based pruning, (iv) a dependence-encoding table with joint-effect semantics, and (v) a specialized training regime driven by nonlinear dependence filtering and random subset training. The literature typically addresses anomaly explanation, feature selection/redundancy control, and representation/reporting as separate modules; it does not teach a unified architecture in which these are coupled so that the output object (a quantitative dependence table) is the product of both the inference flow and the training flow. Combining these is not a predictable variation because each component changes the meaning and stability of the others.
Further, using the median of counterfactual data sets as the comparison point is not an arbitrary baseline. It implicitly targets robustness to skew/outliers and establishes a particular semantics for “difference” that is neither the typical “mean baseline” nor a model-agnostic permutation perturbation. Absent the Applicant’s disclosure, a skilled artisan would not have been led to (a) choose median values from counterfactual sets, and (b) then treat the resulting per-feature numeric differences as the substrate for a dependence-interpretation model that outputs contributions.
The row semantics required by the dependent claims of claim 8 add that each row encode an inference that a target feature changes by an amount due to a joint effect of one or more pluralities of features. Standard importance/explanation outputs are rankings, per-feature attributions, or local explanations of the prediction; they do not require constructing a tabular representation of inferred inter-feature dependence with joint-effect structure and quantified change. Creating that object is a distinct representational goal, and the art does not provide a roadmap for transforming attribution outputs into that specific dependence-table format.
Claims 10–11 recite training the feature-dependence interpretation model using (i) feature-labeling-derived information to reinforce generalization, (ii) minimum-correlated feature subsets, (iii) a nonlinear correlation metric applied pairwise with a threshold-based extraction process, and (iv) random subset training. Even where “reduce redundancy” or “nonlinear dependence” metrics exist, they are typically used for selecting predictors for a downstream predictive model, not as a structured curriculum for training an interpretation model whose outputs are later tabulated as dependence inferences. The claimed regimen therefore is not an obvious optimization; it is a different training objective.
Claim Rejections - 35 USC § 103
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.
Claims 1-3, 5-7, and 14-20 are rejected under 35 U.S.C. 103 as being unpatentable over US Pre-Grant Patent 2021/0056209 (Fox et al; Fox) in view of US Pre-Grant Patent 2022/0245508 (Liao et al; Liao).
Regarding claim 1 and analogous claims 19 and 20:
Fox teaches:
1. An apparatus comprising: a processor; and a computer-readable medium storing instructions which, when executed by the processor, cause the processor to perform first plurality of operations comprised:
(Fox, ¶0022)
“Still another embodiment is a non-transitory computer-readable medium comprising instructions for execution by a computer, the instructions including a computer-implemented method as above, the instructions for implementing the method in a processor [i.e. An apparatus comprising: a processor; and a computer-readable medium storing instructions which, when executed by the processor, cause the processor to perform first plurality of operations comprised:].”
2. detecting an abnormality in a test data set via an abnormality detection model;
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. detecting an abnormality in a test data set via an abnormality detection model;].”
3. generating one or more counterfactual data sets for the test data set
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. detecting an abnormality in a test data set via an abnormality detection model;].”
4. determining a quantitative feature dependence between the test data set and the one or more counterfactual data sets;
(Fox, ¶0135)
“The discriminator network 155 detects signals and scores the events 155b, as discussed elsewhere herein. The events are determined 173 by the discriminator network 155 to be either benign 161 or malicious 163 [i.e. determining a quantitative feature dependence between the test data set and the one or more counterfactual data sets;].”
5. generating, via the abnormality detection model and based on the test data set with the abnormality, a predefined number of counterfactual data sets;
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. generating…a predefined number of counterfactual data sets;].”
5. and wherein the abnormality detection model is trained via a second plurality of operations, the second plurality of operations comprising: generating feature weights based on information comprising feature labelling;
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. generating feature weights based on information comprising feature labelling].”
6. training the abnormality detection model based on the feature weights and training data sets;
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. training the abnormality detection model based on the feature weights and training data sets;].”
7. removing abnormal data from the training data sets;
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions.”
8. and training, after the removing the abnormal data from the training data sets, the abnormality detection model to be a counterfactual generation model that is configured to generate the one or more counterfactual data sets.
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. and training, after the removing the abnormal data from the training data sets, the abnormality detection model to be a counterfactual generation model that is configured to generate the one or more counterfactual data sets].”
Fox does not explicitly teach:
1. determining a causal relationship of the abnormality based on the quantitative feature dependence via a causality discovery algorithm; and generating a causal graph that represents the causal relationship of the abnormality, wherein the causality discovery algorithm is a pairwise causality discovery algorithm, and the determining of the causal relationship of the abnormality based on the quantitative feature dependence comprises; evaluating pairwise causality, via the pairwise causality discovery algorithm.to recognize the causality relationship; and outputting the causality relationship,
Liao teaches:
1. determining a causal relationship of the abnormality based on the quantitative feature dependence via a causality discovery algorithm; and generating a causal graph that represents the causal relationship of the abnormality, wherein the causality discovery algorithm is a pairwise causality discovery algorithm, and the determining of the causal relationship of the abnormality based on the quantitative feature dependence comprises; evaluating pairwise causality, via the pairwise causality discovery algorithm.to recognize the causality relationship;
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm [i.e. determining a causal relationship of the abnormality based on the quantitative feature dependence via a causality discovery algorithm and generating a causal graph that represents the causal relationship of the abnormality, wherein the causality discovery algorithm is a pairwise causality discovery algorithm, and the determining of the causal relationship of the abnormality based on the quantitative feature dependence comprises; evaluating pairwise causality, via the pairwise causality discovery algorithm.to recognize the causality relationship;].”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. A person of ordinary skill in the art seeking to understand what features cause or contribute to Fox’s detected abnormalities, and to generate interpretable explanations, would have been motivated to apply the well-known PC causal-discovery technique of Liao to Fox’s feature/abnormality data. To wit: “Each node of the causal network is a feature of the dataset and each edge in the network represents a causal relationship between the respective pair of connected features as reflected in the labeled dataset. The causal relationships include causal relationships inferred by the model (Liao, ¶0024).”
Regarding claim 2:
Fox and Liao teach the apparatus of claim 1.
Liao teaches:
1. implementing an action to mitigate the abnormality based on the causal graph.
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm.”
Examiner interprets the decision to assign a direction to the abnormal variables in the graph as an attempted action to mitigate under BRI.
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 3:
Fox and Liao teach the apparatus of claim 1.
Fox teaches:
1. acquiring an abnormality score for the test data set via the abnormality detection model;
(Fox, ¶0248)
Abnormal behavior detection scores are generated for human curated and/or for automated decisions. In the implementation of FIG. 8, there are three signals, and a simple heuristic engine is provided wherein the abnormality score is calculated based on how many positive signals are present.
2. determining the abnormality score is greater than a threshold;
(Fox, ¶0251)
“In an embodiment, the system may be set with a threshold abnormality score whereby every abnormality score which exceeds the threshold is deemed to have an unusual risk profile.”
3. and adding, when the abnormality score is greater than the threshold, the test data set to the abnormality detection model, wherein the abnormality detection model is configured to generate the one or more counterfactual data sets.
(Fox, ¶0134)
“A previously-identified set of malicious data may be used again by the generator network 159 to improve simulation data to fool the discriminator network 155a, 155b.”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 5:
Fox and Liao teach the apparatus of claim 1.
Fox teaches:
1. wherein the information comprising feature labelling is input by a user or default information.
(Fox, ¶0135)
“The discriminator network input 155a performs fact discovery on the release events 153, as discussed elsewhere herein. The discriminator network 155 detects signals and scores the events 155b, as discussed elsewhere herein [i.e. wherein the information comprising feature labelling is input by …default information.”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 6:
Fox and Liao teach the apparatus of claim 1.
Fox teaches:
1. wherein the feature labelling comprises: a label of a feature of a data set;
(Fox, ¶0079)
“The adversarial network can feed all of the transactions into the system, and the system figures out what is normal and abnormal; the normal vs. abnormal transactions may be used to determine whether each individual release of a component is normal [i.e. wherein the feature labelling comprises: a label of a feature of a data set].”
2. and a weight for the feature of a data set.
(Fox, ¶0226)
“For a new version of a component, the processor can determined, based on the dataset 665 of release events over time, a historical behavioral analysis of (i) the project to which the component belongs, (ii) historical committer behavior of the committer that committed this version of the component, and/or (iii) historical publisher behavior of the publisher that publishes the component. The historical behavioral analysis may weight most heavily the most recent historical behavior, for example in determining that a potential malicious vector was inserted in the current release code base [i.e. and a weight for the feature of a data set.].”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 7:
Fox and Liao teach the apparatus of claim 1.
Fox teaches:
1. wherein the generating the one or more counterfactual data sets for the test data set is performed via the abnormality detection model that is configured to generate the one or more counterfactual data sets.
(Fox, ¶0065)
“Some embodiments may include an adversarial network, for example, a generative adversarial neural network, in which the system feeds in the normal and abnormal transactions [i.e. wherein the generating the one or more counterfactual data sets for the test data set is performed via the abnormality detection model that is configured to generate the one or more counterfactual data sets].”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 14:
Fox and Liao teach the apparatus of claim 1.
Liao and Fox teach:
1. wherein the determining the causal relationship of the abnormality based on the quantitative feature dependence further comprises: obtaining two features of the abnormality in the test data set; and determining a causal relationship for the two features does not exist in information comprising feature dependence;
(Liao, ¶0099)
“Deletion of an edge P to Q in the causal network signify the intent of the SME 520 of making variables P and Q statistically independent from each other [i.e. wherein the determining the causal relationship of the abnormality based on the quantitative feature dependence further comprises: obtaining two features of the abnormality in the test data set; and determining a causal relationship for the two features does not exist in information comprising feature dependence;].”
2. generating, via the abnormality detection model and based on the test data set with the abnormality, a predefined number of counterfactual datasets;
(Fox, ¶0135)
“The discriminator network 155 detects signals and scores the events 155b, as discussed elsewhere herein. The events are determined 173 by the discriminator network 155 to be either benign 161 or malicious 163 [i.e. generating, via the abnormality detection model and based on the test data set with the abnormality, a predefined number of counterfactual datasets;].”
3. recognize a causality relationship;
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm [i.e. recognize a causality relationship;].”
4. and outputting the causality relationship.
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm [i.e. and outputting the causality relationship].”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 15:
Fox and Liao teach the apparatus of claim 1.
Liao teaches:
1. wherein the causality relationship is recognized after iteration of possible feature pairs of a subset of elements in a table that quantitatively describes feature dependence.
(Liao, ¶0102)
“With reference to FIG. 8, this figure depicts a graph of data used for modifying a query strategy based on a modification instruction to add or strengthen an edge from an SME 520 in accordance with an illustrative embodiment. Specifically, arrow A3 shows strengthening an edge by increasing the slope of a distribution to F5 from that of F6a, and arrow A4 shows adding an edge by increasing the slope of a distribution to F5 from that of F6b [i.e. wherein the causality relationship is recognized after iteration of possible feature pair].”
(Liao, ¶0103)
“Adding an edge P to Q between two previously unconnected nodes P and Q signify the intent of SME 520 to establish statistical dependence between two independent variables. Similarly, strengthening an edge represents having a stronger dependence relation than before [i.e. of a subset of elements in a table that quantitatively describes feature dependence.]”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 16:
Fox and Liao teach the apparatus of claim 1.
Liao teaches:
1. wherein the pairwise causality discovery algorithm comprises an additive-noise model (ANM) or an information geometric causal inference (IGCI) model.
(Liao, ¶0088)
“The SME 520 then corrects the causal network by modifying the associated edge, for example by adding, strengthening, deleting, or diluting the edge. Such intervention will change the query strategy such that it will aim to neutralize the unintended causal relationships while also considering the performance goal [i.e. wherein the pairwise causality discovery algorithm comprises an additive-noise model (ANM)].”
Examiner notes that edge manipulation can be thought of as adding noise, and thus incorporating an additive noise model under BRI.
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 17:
Fox and Liao teach the apparatus of claim 1.
Liao teaches:
1. wherein the determining the causal relationship of the abnormality based on the quantitative feature dependence comprises: obtaining a plurality of features of the abnormality in the test data set;
(Liao, ¶0077)
“In the illustrated embodiment, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm. Each node of the causal network is a feature of the dataset and each edge in the network represents a causal relationship between the respective pair of connected features as reflected in the dataset already labeled [i.e. obtaining a plurality of features of the abnormality in the test data set;].”
2. determining a causal relationship for the plurality of features exists in information comprising feature dependence;
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm [i.e. determining a causal relationship for the plurality of features exists in information comprising feature dependence;].”
3. and outputting the causal relationship.
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm [i.e. and outputting the causality relationship].”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
Regarding claim 18:
Fox and Liao teach the apparatus of claim 1.
Liao teaches:
1. wherein the causal relationship represents the quantitative feature dependence of the abnormality.
(Liao, ¶0024)
“In exemplary embodiments, the active learning training cycle periodically generates a causal network from the labeled data using a causal discovery algorithm, such as the Peter and Clark (PC) algorithm.”
One of ordinary skill in the art, at the time the invention was filed, would have been motivated to modify Fox with Liao. The motivation is the same as claim 1.
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129