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 § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-12 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1,
Step 1:
Claim 1 recites a method, one of the four statutory categories of patentable subject matter.
Step 2A, Prong I:
Claim 1 further recites the limitations of:
“generating, ..., a multi-dimensional perturbed data by applying constrained perturbations on a multi-dimensional tabular base data comprising a plurality of categorical features and a plurality of continuous features, wherein the constrained perturbations generate the multi-dimensional perturbed data in vicinity of a distribution of the multi-dimensional tabular base data” The limitation recites a mental process. A person can mentally or manually modify original table data while keeping them close to the original data pattern or original data category. For example, given a table of data with label, a person can manually change the value or category within the table of data to perturb table data in a constrained manner.
“applying, ..., a non-linear dimensionality reduction technique on the multi-dimensional perturbed data and the multi-dimensional base data to generate a dimensionality reduced perturbed data and a dimensionality reduced base data” The limitation recites a mental process, as well as a mathematical concept. The non-linear dimensionality reduction technique is recited to be a t-distributed Stochastic Neighbor Embedding (t-SNE) technique, which involves mathematical algorithm that can be mentally performed.
“... using a first local maxima of a Silhouette score technique to identify a plurality of main clusters of the dimensionality reduced base data, wherein the plurality of main clusters capture the distribution of the dimensionality reduced base data and are identified as an optimal number of clusters”. The limitation recites a mental process, as well as a mathematical concept. The Silhouette score technique involves mathematical algorithm that can be mentally performed. Furthermore, a person can mentally calculate the Silhouette score to test different cluster options and choosing a good number to group data, thereby determining a good data group that correspond with each other. A person can mentally evaluate whether the data are close with each other to group them together, especially via using mathematical concept such as the Silhouette score.
“selecting, ..., a subset of perturbed data samples from among the dimensionality reduced perturbed data that lie within a median cluster distance from a cluster center of a closest cluster among the optimal number of clusters”. The limitation recites a mental process, as well as a mathematical concept. The limitation recites calculating distance between center and data point, which involves mathematical algorithm that can be mentally performed. Furthermore, a person can mentally determine from such distance calculation whether to keep or drop data that is too far from the center point. Such evaluation involves mathematical concept and mental process.
“generating, ..., tabular synthetic data having the distribution within the distribution of the multi-dimensional base data by selecting the multi-dimensional perturbed data, associated with the dimensionality reduced perturbed data lying within the median cluster distance”. The limitation recites a mental process, as well as a mathematical concept. The limitation recites calculating distance between center and data point, and further evaluate data that is close to the center of the nearest group, which involves mathematical algorithm to calculate distance that can be mentally performed, as well as mental evaluation of the distance to keep or drop data.
Step 2A, Prong II:
Claim 1 recites the following additional elements:
“... via the one or more hardware processors... ” This additional element is a high-level recitation of generic computer components used as a tool, and does not provide integration into a practical application.
“training, ..., a plurality of Gaussian Mixture Models (GMMs) on the dimensionality reduced base data ...” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the training of plurality of Gaussian Mixture models on the dimensionality reduced data without providing the technical detail or an unconventional training algorithm, or improvement over a computer hardware. The limitation simply recites conventional black-box usage of machine learning training and thus, the additional element does not provide integration into a practical application.
Step 2B:
When considered individually or in combination, the additional limitations and elements of claim 1 does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application. The additional elements of outlined in Step 2A performing functions as designed simply accomplishes execution of the abstract ideas.
The additional element “... via the one or more hardware processors...” is a high-level recitation of generic computer components used as a tool, and does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application.
The additional element “training, ..., a plurality of Gaussian Mixture Models (GMMs) on the dimensionality reduced base data ...” recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not amount to significantly more than the judicial exception for the same reasons discussed above.
In conclusions from above for the elements considered as a mental process/mathematical concept, elements reciting high-level recitation of generic computer components used as a tool, and elements reciting a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f) are carried over and do not provide significantly more than the abstract idea. Looking at the limitations in combination and the claims as a whole does not change this conclusion and the claim is ineligible.
Therefore, additional limitations of claim 1 do not amount to significantly more than the judicial exception.
Thus, claim 1 recites abstract ideas with additional elements rendered at a high level of generality resulting in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception.
Therefore, claim 1 is not patent eligible.
Regarding claim 2 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 2 recites the element:
“The method of claim 1, wherein the tabular synthetic data is processed to generate labelled training data for building Machine Learning (ML) models” This limitation recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the application of modified table data (tabular synthetic data) to generated labelled data to train machine learning model. However, training a machine learning model with labelled data or modified/perturbed data is a conventional practice in machine learning or black-box application of machine learning model and thus, the additional element does not provide integration into a practical application.
Thus, claim 2 recites additional elements rendered at a high level of generality resulting in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 2 is not patent eligible.
Regarding claim 3 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 3 recites the elements:
“The method of claim 1, wherein the constrained perturbations applied on the plurality of continuous features are based on a Coefficient of Variation (CV) score for each feature obtained from distribution of a percentage of sample data selected from among the multi-dimensional tabular base data” The limitation recites a mental process, and a mathematical concept. The limitation recites a Coefficient of Variation (CV) score, which involves mathematical algorithm that can be performed by a human mind.
“wherein the constrained perturbations applied on the plurality of categorical features are obtained by random sampling from set of feature values of a sample data such that it covers 90% of the percentage of sample data selected from among the multi-dimensional tabular base data” The limitation recites a mental process. The limitation recites a random sampling of value at it covers 90% of the percentage of sample data selected, which involves a data modification process that can be manually perform.
Thus, claim 3 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 3 is not patent eligible.
Regarding claim 4 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 4 recites the elements:
“The method of claim 1, wherein the non-linear dimensionality reduction technique is t-distributed stochastic neighbor embedding (t-SNE)” The limitation recites a mental process, as well as a mathematical concept. The non-linear dimensionality reduction technique is recited to be a t-distributed Stochastic Neighbor Embedding (t-SNE) technique, which involves mathematical algorithm that can be mentally performed.
Thus, claim 4 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 4 is not patent eligible.
Regarding claim 5 which recites a system, one of the four statutory categories of patentable subject matter.
Claim 5 recites the elements:
“a memory storing instructions; one or more Input/Output (I/O) interfaces; and one or more hardware processors coupled to the memory via the one or more I/O interfaces, wherein the one or more hardware processors are configured by the instructions”
The limitation is a high-level recitation of generic computer components used as a tool, and does not amount to significantly more than the judicial exception and does not integrate the abstract idea into a practical application.
The applicant is further directed to the rejection of claim 1 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 6 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 2 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 7 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 3 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 8 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 4 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 9 which recites a machine, one of the four statutory categories of patentable subject matter.
Claim 9 recites the elements:
“One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors”
The limitation is a high-level recitation of generic computer components used as a tool, and does not amount to significantly more than the judicial exception and does not integrate the abstract idea into a practical application.
The applicant is further directed to the rejection of claim 1 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 10 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 2 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 11 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 3 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 12 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 4 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
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.
Claim(s) 1, 2, 4-6, 8-10, 12 are rejected under 35 U.S.C. 103 as being unpatentable over Nia et.al (US 20220129791 A1) in view of Njie et.al (US 20190347567 A1), further in view of Lavorini et.al (NPL: Gaussian Mixture Model clustering: how to select the number of components (clusters)), further in view of Avidan et.al (US 8363961 B1)
Regarding claim 1,
Nia teaches or at least suggest the limitation “generating, via one or more hardware processors, a multi-dimensional perturbed data by applying constrained perturbations on a multi-dimensional tabular base data comprising a plurality of categorical features and a plurality of continuous features, wherein the constrained perturbations generate the multi-dimensional perturbed data in vicinity of a distribution of the multi-dimensional tabular base data” (paragraph 53 “Techniques are described herein in the context of tabular data, which comprises values with associated labels, such as column-based data, key/value pair-based data, or attribute and attribute value-formatted data”, paragraph 65 “Sx is expressed by an n-dimensional hypersphere of radius”, paragraph 78 “Accordingly, SEA 110 identifies and/or generates data samples within the defined hypersphere that: (a) preserve the training data set characteristics, such as feature correlation and distribution; (b) cover black-box complexities affecting the model output in the vicinity of the data sample; and (c) are highly local to the data sample of interest”, paragraph 77 “the generated samples are guaranteed to be in close proximity to the target data sample (i.e., within the hypersphere)” paragraph 80-84 “Generating new data samples: a. For numerical features ... minor perturbations of feature values (xi) to increase the variety of data samples within the data sample feature space ... bounding of the perturbation with relation to the size of the data sample feature space being explored ...”, and paragraph 85-86 “b. For categorical features ... Generate random values based on the frequency of feature values in the local training points obtained in step one” Nia discloses a systematic an approach for explaining machine learning predictions. Within the disclosure, Nia discloses technique to generate local data samples around a given target data sample, which improves on exhaustive or random data sample generation algorithms by identifying a hypersphere (or data sample neighborhood) to generate data sample that closely relate to original data. The generation of data sample is based upon original tabular training data. The n-dimensional hypersphere of data points suggests multi-dimensional tabular data. The technique to generate data sample involves perturbation based on the radius of the hypersphere of data points, which corresponds to the constrained perturbation as claimed, because the perturbation is required to follow a set radius of the hypersphere and the number of features, thus the amount of perturbation is controlled and not over-sized or under-sized, which suggests a constraint technique. Furthermore, the tabular data comprises of numerical features and categorical features, which corresponds to the categorical features and continuous features, as claimed. Accordingly, because Nia’s new samples are bounded within a hypersphere, generated from local training-point feature distributions/frequencies, and preserve feature correlation and distribution, the new samples correspond to the multi-dimensional perturbed data in vicinity of a distribution of the multi-dimensional tabular base data, as claimed.)
Nia does not teach the limitation “applying, via the one or more hardware processors, a non-linear dimensionality reduction technique on the multi-dimensional perturbed data and the multi-dimensional base data to generate a dimensionality reduced perturbed data and a dimensionality reduced base data”. However, Njie teaches or at least suggest the limitation (paragraph 53 ““t-Distributed Stochastic Neighbor Embedding” or “t-SNE” is a nonlinear, AI-based dimension reduction method that allows for visualization of high dimensional data by giving each data point a location in a map, e.g., a two or three-dimensional map”, and paragraph 56 “the methods of the present disclosure use an inverse approach, first applying the nonlinear dimension reduction system (e.g., t-SNE) and subsequently using the output of this method as the input for training of an artificial neural network” Njie discloses AI-based dimension reduction method that allows for visualization of high dimensional data by giving each data point a location in a map, wherein one of ordinary skill in the art would have been motivated to perform the AI-based dimension reduction upon Nia’s new samples and original training data because Nia’s new samples and original training data are both multi-dimensional tabular data points in the same feature space. Applying, Njie’s dimension reduction technique to both datasets would place them in a common lower dimensional space, thereby allowing the generated new samples to be compared with the original training data distribution for subsequent processing.)
Before the effective filing date, it would have been obvious to one of ordinary skill in the art to combine the teaching of generating new samples based on original training tabular data based on perturbation by Nia with the teaching of applying nonlinear dimension reduction of data technique by Njie. The motivation to combine the teaching is referred to in Njie’s disclosure (paragraph 73 “The aim of dimension reduction is to transform high dimensional problems into a visually tractable form. For instance, in genetics, the Homo sapiens genome is composed of some 3 billion characters that in the imagination of many geneticists form a linear string. A non-linear dimension reduction viewpoint would rephrase this as at least three billion dimensions, and since humans have difficulty making sense of more than three dimensions, reduction of more than three billion dimensions to two or three is a sensible means to more insightfully understand the genome” Njie discloses the benefit of the dimension reduction technique, which is to transform high dimensional problems into a visually tractable form to gain insightfully understanding of data. One of ordinary skill in the art would have been motivated to perform the AI-based dimension reduction upon Nia’s new samples and original training data because Nia’s new samples and original training data are both multi-dimensional tabular data points in the same feature space. Applying, Njie’s dimension reduction technique to both datasets would place them in a common lower dimensional space and obtain the improvement of a visually tractable form that allow insightfully understanding of data as suggested by Njie.)
Nia/Njie does not teach the limitation “training, via the one or more hardware processors, a plurality of Gaussian Mixture Models (GMMs) on the dimensionality reduced base data using a first local maxima of a Silhouette score technique to identify a plurality of main clusters of the dimensionality reduced base data, wherein the plurality of main clusters capture the distribution of the dimensionality reduced base data and are identified as an optimal number of clusters”. However, Lavorini teaches or at least suggest the limitation (Page 1 “the GMM as a k-means which is able to form stretched clusters, like the ones you can see in Figure 2”, Page 3 “... run the fitting procedure several times, and consider the mean value and the standard deviation for each configuration”, Page 3-4 “... Cluster performance evaluation(s) Since we don’t know the ground truth of our cluster generators, i.e. we are not aware of the original distribution which generated the data ... Silhouette score This score, as clearly stated by the SKLearn developers, consider two measures: The mean distance between a sample and all other points in the same cluster. The mean distance between a sample and all other points in the next nearest cluster ...”, Page 7 “We can say that the good configuration, which takes in account both of the amount of information included (=biggest possible number of clusters) and on the stability of the fitting procedure (=lowest possible GMMs distance), is the one which considers six cluster” Lavorini discloses fitting/training Gaussian Mixture Model to identify clusters and using Silhouette score to evaluate Gaussian model clusters. Within the disclosure, Lavorini discloses using Gaussian Mixture Models (GMMs) for clustering data, where each GMM configuration corresponds to a different number of Gaussian components/clusters. Lavorini explains that, when the ground-truth number of clusters or Gaussian components is unknown, an automated method is needed to determine the “right” number of clusters. Lavorini further discloses evaluating the GMM cluster configurations using a Silhouette score, thereby checking whether the clusters are compact and well separated. Thus, Lavorini teaches or suggests training a plurality of GMMs with different numbers of Gaussian components/clusters and using a Silhouette score technique to identify an appropriate/optimal number of clusters. The identified GMM components correspond to the claimed main clusters because the Gaussian components define clusters of the data. Although Lavorini does not expressly recite “first local maxima,” Lavorini teaches evaluating Silhouette scores across different numbers of clusters to select the right number of clusters. A person of ordinary skill in the art would have understood that selecting a local peak, including a first local maximum, of the Silhouette-score curve is a predictable implementation of Lavorini’s Silhouette-based selection because the local peak indicates a cluster number where compactness and separation improve relative to neighboring cluster numbers. A person of ordinary skill in the art would have understood that the selected GMM clusters capture the distribution of Nia’s reduced original training data because Lavorini’s GMM clusters model the data using Gaussian components and are selected based on compactness and separation. Therefore, applying Lavorini’s GMM/Silhouette technique to the reduced original training data would identify clusters representing the distribution/cluster structure of the original training data for later comparison with Nia’s reduced generated samples.)
Before the effective filing date, it would have been obvious to one of ordinary skill in the art to combine the teaching of the teaching of generating new samples based on original training tabular data based on perturbation by Nia, the teaching of applying nonlinear dimension reduction of data technique by Njie, with the teaching of Gaussian Mixture Model clustering and using Silhouette score to evaluate Gaussian model clusters by Lavorini. The motivation to combine the teaching is referred to in Lavorini’s disclosure (Page 2 “... you want to discern how many clusters we have (or, if you prefer, how many gaussians components generated the data), and you don’t have information about the “ground truth” ... here we want to check for an automated method for finding the “right” number of clusters ...”, Page 3-5 “The easiest way of dealing with it is to run the fitting procedure several times, and consider the mean value and the standard deviation for each configuration ... Silhoutte score ... i.e. it checks how much the clusters are compact and well separated. The more the score is near to one, the better the clustering is.” Lavorini discloses when the ground-truth number of clusters/Gaussian components is unknown, an automated technique is needed to determine the right number of clusters. Lavorini further recognizes that GMM fitting may vary due to the EM algorithm converging to local optima, and therefore teaches running GMM fitting multiple times for each configuration and evaluating the configurations using Silhouette score to identify compact and well-separated clusters. A person of ordinary skill in the art would have been motivated to apply Lavorini’s technique to the dimensionality reduced data generated from Nia’s original training data in view of Njie, because Njie’s t-SNE output provides lower-dimensional data points suitable for clustering, and Lavorini teaches an automated technique for identifying the appropriate number of clusters based on Gaussian Mixture model when the true cluster structure is unknown. Applying Lavorini’s technique to the reduced original training data would identify compact and well-separated clusters that represent the distribution of Nia’s original training data. These clusters would then provide a reference distribution to comparing Nia’s new samples and determining whether the new samples remain close to the original training data distribution. Therefore, the teaching combination would have predictably improved Nia’s new generated samples so that they could be compared against the original data distribution and selected with better reliability.)
Nia/Njie/Lavorini does not teach the limitation “selecting, via the one or more hardware processors, a subset of perturbed data samples from among the dimensionality reduced perturbed data that lie within a median cluster distance from a cluster center of a closest cluster among the optimal number of clusters”. However, Avidan teaches or at least suggest the limitation (Column 1 lines 52-62 “for a point p selected from within set P, identifying a set S including a plurality of points within set P that neighbor point p according to a distance metric defined on the multidimensional space; (b) generating a centroid point c of set S, wherein centroid point c is a member of set S; (c) determining whether a distance between the centroid point c and the point p satisfies a threshold value; (d) in response to determining that the distance between the centroid point c and the point p does satisfy the threshold value, returning centroid point c as a mode m of set P and terminating”, Column 4 lines 26-28 “In various embodiments, the neighborhood of the given point may be defined as a region that contains a specific number of data points k that are closest to the given point”, Column 9 lines 1-6 “In one embodiment, the threshold value may be a numerical value such as zero or a nonzero value, and determining whether the distance satisfies the threshold value may include determining whether the distance is less than (or less than or equal to) the threshold value”, and Column 9 lines 13-22 “In one embodiment, the centroid-shift mode-seeking algorithm may be repeatedly applied using each point p within data set P ... as the initial point. Once the mode has been identified for each of multiple points p, those points that converge to a common mode (either exactly, or in some cases within a threshold of error) may be identified as members of the same cluster of points. That is, the mode returned by the centroid-shift algorithm may be used as a criterion for cluster membership” Avidan discloses selecting or evaluating a multidimensional data point based on its distance from a centroid or central point. Avidan further teaches that the neighboring points may be the closest points to a given point and that points converging to a common mode may be identified as members of the same cluster. This corresponds to the claimed selection step because the claim similarly evaluates whether a data sample is close enough to a cluster center of the closest cluster before selecting that sample. In both Avidan and the claim, a data point is evaluated relative to a nearby cluster or centroid region, and the point is selected or treated as belonging to that region when its distance from the centroid or cluster center satisfies a threshold. Avidan’s set S corresponds to or suggests a local cluster/neighborhood region because S includes points neighboring point p according to a distance metric, and Avidan generates centroid c from S for determining whether point p belongs with that region. Furthermore, in view of the combination with Njie that teaches the reduce data dimensional, and Lavorini provides the GMM clusters and cluster centers from Nia’s reduced original training data, accordingly, Avidan’s point p may correspond to one of Nia’s reduced generated or perturbed samples, and Avidan’s centroid c corresponds to the center of the GMM cluster as identified by Lavorini. Although Avidan does not expressly recite “median cluster distance,” Avidan teaches using a point-to-centroid distance threshold for cluster membership or selection. A person of ordinary skill in the art would have been motivated to use the median cluster distance as the threshold because it represents a typical point-to-center distance within the cluster and is less affected by outliers. Therefore, the combination teaches or suggests selecting reduced generated or perturbed samples that lie within a median cluster distance from the closest cluster center, as claimed.)
Before the effective filing date, it would have been obvious to one of ordinary skill in the art to combine the teaching of the teaching of generating new samples based on original training tabular data based on perturbation by Nia, the teaching of applying nonlinear dimension reduction of data technique by Njie, the teaching of Gaussian Mixture Model clustering and using Silhouette score to evaluate Gaussian model clusters by Lavorini, with the teaching of selecting data points based on distance from a cluster centroid by Avidan. The motivation to combine the teaching is referred to in Avidan’s disclosure (Column 1 lines 35-40 “To be effective in the context of imaging, clustering algorithms thus need to readily extend to highly dimensional data. However, many existing clustering techniques are not robust, in that they are sensitive to noise in the image data and/or small variations in the parameters that govern the algorithm's performance. Moreover, complex images may include vast numbers of data points of high dimensionality, and many existing clustering techniques fail to scale well as the data set size increases”, and Column 9 lines 13-22 “In one embodiment, the centroid-shift mode-seeking algorithm may be repeatedly applied using each point p within data set P ... as the initial point. ... those points that converge to a common mode ... may be identified as members of the same cluster of points. That is, the mode returned by the centroid-shift algorithm may be used as a criterion for cluster membership” Avidan discloses the extension of clustering algorithms to highly dimensional data via a centroid-shift mode-seeking algorithm to improve the current clustering technique. The technique by Avidan allows for the evaluation of distance from one or more data points to a centroid of cluster, which helps determine whether a data point belongs to or corresponds to a cluster region containing data points in the same vicinity. A person of ordinary skill in the art would have been motived to apply Avidan’s distance-to-centroid evaluation to the reduced generated samples of Nia/Njie/Lavorini because the reduced training data clusters identified by Lavorini provide cluster centers representing the original training data distribution. Avidan’s technique would allow each reduced generated sample to ve evaluated relative to a closest cluster centroid using a distance threshold, thereby determining whether the generated sample remains within the same distribution as the original training-data clusters. This would have predictably improved the combination by providing a cluster-based filtering/selection step that retains generated samples close to the original training data distribution and rejectes samples that does not meet the distance threshold. Thus, incorporating the teaching of Avidan would have predictably improved the teaching combination in evaluating the generated new samples data and the original training data in Nia in view of Njie and Lavorini.)
Nia in view of the teachings of Njie/Lavorini/Avidan above teaches or at least suggest the limitation “generating, via the one or more hardware processors, tabular synthetic data having the distribution within the distribution of the multi-dimensional base data by selecting the multi-dimensional perturbed data, associated with the dimensionality reduced perturbed data lying within the median cluster distance” (Nia discloses the generating of new sample data (perturb data) at paragraph 53, 65 and 78 as recited above. Njie discloses the “t-Distributed Stochastic Neighbor Embedding” or “t-SNE” nonlinear dimension reduction method as recited at paragraph 53 above. Lavorini discloses the generating of multiple clusters via Gausian mixture models and evaluating using Silhouette score as recited above. Finally, Avidan discloses at Column 9 lines 13-22 “In one embodiment, the centroid-shift mode-seeking algorithm may be repeatedly applied using each point p within data set P ... as the initial point. Once the mode has been identified for each of multiple points p, those points that converge to a common mode (either exactly, or in some cases within a threshold of error) may be identified as members of the same cluster of points. That is, the mode returned by the centroid-shift algorithm may be used as a criterion for cluster membership”. Nia discloses generating new tabular data samples from original training data, where the generated samples preserve training-data characteristics, such as feature correlation and distribution, and remain local to the target data sample within a defined hypersphere. Thus, Nia teaches generated/perturbed tabular samples that remain close to the original tabular training-data distribution. Although Nia does not expressly teach further selecting the generated/perturbed tabular samples based on reduced-space cluster distance, the other teachings in the combination suggest such selection. In particular, Njie teaches applying nonlinear dimensionality reduction to high-dimensional data, which would place Nia’s original training data and generated samples in a common reduced space. Lavorini teaches applying GMM/Silhouette clustering to identify clusters representing the distribution of the reduced original training data. Avidan teaches evaluating a data point based on its distance from a centroid/cluster center and determining whether the distance satisfies a threshold for cluster membership. Accordingly, the combined teachings suggest selecting Nia’s generated/perturbed tabular samples by evaluating their corresponding reduced samples against the closest cluster center of the reduced original training data and selecting the corresponding original generated/perturbed tabular samples when the reduced samples lie within the median cluster distance. A person of ordinary skill in the art would have been motivated to make this selection because it would retain generated samples that remain within the original training-data distribution and reject samples that are too far from the learned cluster regions, thereby generating new tabular data samples (synthetic data) having a distribution within the distribution of the original tabular training data.)
Regarding claim 2 depends on claim 1, thus the rejection of claim 1 is incorporated.
Nia teaches or at least suggest the limitation “The method of claim 1, wherein the tabular synthetic data is processed to generate labelled training data for building Machine Learning (ML) models” (paragraph 57 “At step 206, labels for the generated data samples are computed using black-box ML model 102 ... Specifically, the predictions of black-box ML model 102 are used as the expected predictions for surrogate training data set 126 being produced”, and paragraph 154 “Model training may be supervised or unsupervised. For supervised training, the desired (i.e., correct) output is already known for each example in a training set. The training set is configured in advance by (e.g., a human expert, or via the labeling algorithm described above) assigning a categorization label to each example” Nia discloses the generated new data samples are used for training the ML model, wherein labels are generated for the generated data samples, indicating a labelled generated data samples for training ML models, which corresponds to or at least suggest the processing of tabular synthetic data to generate labelled training data for building Machine Learning (ML) models, as claimed.)
Regarding claim 4 depends on claim 1, thus the rejection of claim 1 is incorporated.
Njie teaches the limitation “The method of claim 1, wherein the non-linear dimensionality reduction technique is t-distributed stochastic neighbor embedding (t-SNE)” (paragraph 53 ““t-Distributed Stochastic Neighbor Embedding” or “t-SNE” is a nonlinear, AI-based dimension reduction method that allows for visualization of high dimensional data by giving each data point a location in a map, e.g., a two or three-dimensional map”, and paragraph 56 “the methods of the present disclosure use an inverse approach, first applying the nonlinear dimension reduction system (e.g., t-SNE) and subsequently using the output of this method as the input for training of an artificial neural network” Njie discloses AI-based dimension reduction method that allows for visualization of high dimensional data by giving each data point a location in a map, which is “t-Distributed Stochastic Neighbor Embedding” or “t-SNE”, which corresponds to the non-linear dimensionality reduction technique, as claimed.)
Regarding claim 5,
Nia teaches the limitation “a memory storing instructions; one or more Input/Output (I/O) interfaces; and one or more hardware processors coupled to the memory via the one or more I/O interfaces, wherein the one or more hardware processors are configured by the instructions” (paragraph 170 “one or more general purpose hardware processors programmed to perform the techniques pursuant to program instructions in firmware, memory, other storage, or a combination”, paragraph 185 “The OS 910 manages low-level aspects of computer operation, including managing execution of processes, memory allocation, file input and output (I/O), and device I/O”, paragraph 186 “Software system 900 includes a graphical user interface (GUI) 915, for receiving user commands and data ... These inputs, in turn, may be acted upon by the system 900 in accordance with instructions from operating system 910 and/or application(s) 902” Nia discloses memory to store program instructions to be performed by hardware processors, wherein the processing include an input/output device and interface to implement the instruction.)
The applicant is further directed to the rejection of claim 1 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 6 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 2 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 8 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 4 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 9,
Nia teaches the limitation “One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors” (paragraph 176 “The term “storage media” as used herein refers to any non-transitory media that store data and/or instructions that cause a machine to operate in a specific fashion. Such storage media may comprise non-volatile media and/or volatile media” Nia discloses non-transitory media that store data information and instructions that cause a machine to operate.)
The applicant is further directed to the rejection of claim 1 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 10 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 2 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 12 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 4 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Claim(s) 3, 7, 11 are rejected under 35 U.S.C. 103 as being unpatentable over Nia et.al (US 20220129791 A1) in view of Njie et.al (US 20190347567 A1), further in view of Lavorini et.al (NPL: Gaussian Mixture Model clustering: how to select the number of components (clusters)), further in view of Avidan et.al (US 8363961 B1), further in view of Frost et.al (NPL: Coefficient of Variation in Statistics)
Regarding claim 3 depends on claim 1, thus the rejection of claim 1 is incorporated.
Nia teaches or at least suggest the limitation “wherein the constrained perturbations applied on the plurality of categorical features are obtained by random sampling from set of feature values of a sample data such that it covers 90% of the percentage of sample data selected from among the multi-dimensional tabular base data” (paragraph 85-86 “For categorical features: Generate random values based on the frequency of feature values in the local training points obtained in step one. Using the frequency of feature values in the training data samples that fall within sx ensures that the randomly generated features are also within sx” Nia discloses that random values are generated based on the frequency of feature values in the local training points for categorical features. Nia further discloses that using the frequency of feature values in the training data samples that fall within sx ensures that the randomly generated features are also within sx. Nia therefore teaches or suggests generating categorical feature perturbations by randomly sampling categorical values from feature values of local training/sample data, wherein the sampling is constrained by the observed frequency distribution of those feature values. Nia’s local training points within sx correspond to sample data selected from the tabular base data, and Nia’s frequency-based random value generation corresponds to random sampling from the set of categorical feature values of that sample data. Although Nia does not expressly recite “90%,” a person of ordinary skill in the art would have found it obvious to use a high-coverage threshold, such as 90%, so that the sampled categorical values remain representative of the selected sample-data distribution while avoiding rare or outlier values. Such a threshold is a predictable implementation of Nia’s teaching to generate categorical values based on the frequency of feature values in the local training points and to keep the generated features within sx.
Nia/Njie/Lavorini/Avidan does not teach the limitation “The method of claim 1, wherein the constrained perturbations applied on the plurality of continuous features are based on a Coefficient of Variation (CV) score for each feature obtained from distribution of a percentage of sample data selected from among the multi-dimensional tabular base data”. However, Frost teaches or at least suggest this limitation (Page 1-2 “How to Calculate the Coefficient of Variation Calculating the coefficient of variation involves a simple ratio. Simply take the standard deviation and divide it by the mean ... In general, higher values represent a greater degree of relative variability ...”, Page 3-4 “When you measure a characteristic that has a wide range of values, you’d often expect the mean and standard deviation to change together. This phenomenon frequently occurs in cross-sectional data. In these cases, you want to know how the standard deviation compares relatively to the vastly different means ... Analysts frequently use the coefficient of variability when their dataset has a broad range of means, as shown in the previous example ... The coefficient of variation is particularly helpful when your data follow a lognormal distribution. In these distributions, the standard deviation changes depending on the portion of the distribution you are assessing. However, the coefficient of variation remains constant throughout a lognormal distribution” Frost discloses that the coefficient of variation (CV) is calculated by taking the standard deviation and dividing it by the mean. Frost further discloses that the CV is a relative measure of variability, where higher CV values indicate a greater degree of relative variability, and that CV is useful for comparing variability between characteristics or datasets having different means.This maps to the claimed CV score because the claimed continuous features are numerical characteristics of the tabular data, and Frost’s CV provides a feature-level measure of how much each continuous feature varies relative to its mean. In view of Nia, the combination teaches the claimed constrained perturbation because Nia teaches generating new samples for numerical features using minor perturbations of feature values, where the perturbation amount is bounded based on the feature space being explored. A person of ordinary skill in the art would have been motivated to use Frost’s CV score for to control or scale Nia’s numerical-feature perturbation because CV indicates the relative variability of each feature. Thus, features having greater relative variability could be perturbed within a larger range, while features having lower relative variability could be perturbed within a smaller range. Accordingly, Nia in view of Frost teaches or suggests constrained perturbations applied to continuous features based on a CV score for each feature obtained from the distribution of selected sample data from the tabular base data.)
Before the effective filing date, it would have been obvious to one of ordinary skill in the art to combine the teaching of the teaching of generating new samples based on original training tabular data based on perturbation by Nia, the teaching of applying nonlinear dimension reduction of data technique by Njie, the teaching of Gaussian Mixture Model clustering and using Silhouette score to evaluate Gaussian model clusters by Lavorini, and the teaching of selecting data points based on distance from a cluster centroid by Avidan with the teaching of coefficient of variation in statistics by Frost. The motivation to combine the teaching is referred to in Frost’s disclosure (Page 3-4 “Use the coefficient of variation when you want to compare variability between: Groups that have means of very different magnitudes. Characteristics that use different units of measurements ... When you measure a characteristic that has a wide range of values, you’d often expect the mean and standard deviation to change together. This phenomenon frequently occurs in cross-sectional data. In these cases, you want to know how the standard deviation compares relatively to the vastly different means ... Analysts frequently use the coefficient of variability when their dataset has a broad range of means, as shown in the previous example”. Frost discloses that the coefficient of variation is useful for comparing variability between characteristics having different means, different units, or wide ranges of values. Frost explains that CV normalizes variability by comparing the standard deviation relative to the mean, so variability can be compared across different characteristics on a common relative basis. In view of Frost, a person of ordinary skill in the art would have been motivated to apply CV to Nia’s numerical/continuous tabular features because Nia’s tabular data includes multiple features that may have different scales, means, and distributions. Nia already teaches generating new samples by applying minor perturbations to numerical feature values and bounding the perturbation amount based on the feature space being explored. Using Frost’s CV score for each continuous feature would improve Nia’s perturbation technique by allowing the perturbation amount to be scaled according to each feature’s relative variability, rather than applying the same absolute perturbation to features with different ranges or units. This would keep the generated samples consistent with the distribution of the selected original training data.)
Regarding claim 7 depends on claim 5, thus the rejection of claim 5 is incorporated. The applicant is further directed to the rejection of claim 3 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Regarding claim 11 depends on claim 9, thus the rejection of claim 9 is incorporated. The applicant is further directed to the rejection of claim 3 above because the claim recites similar limitations and processing steps, thus the claim is rejected under the same rationale.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DUY TU DIEP whose telephone number is (703)756-1738. The examiner can normally be reached M-F 8-4:30.
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.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached at (571) 270-3428. 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.
/DUY T DIEP/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights