DETAILED ACTION
This action is responsive to the amendment filed on 07/15/2026. Claims 1-20 are pending in the case. Claims 1-2, 4-5, 7, 12, 14-15, and 17 are currently amended. Claims 1, 2, and 12 are independent claims.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 103
The 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 1-2, 10-12, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Shuqair et al., Incremental Learning in Time-series Data using Reinforcement Learning, 2022 IEEE International Conference on Data Mining Workshops (ICDMW), Orlando, FL, USA, 2022, pp. 868-875, doi: 10.1109/ICDMW58026.2022.00115, hereinafter referred to as “Shuqair” in view of Liu et al., An Ensemble of Classifiers Based on Positive and Unlabeled Data in One-Class Remote Sensing Classification, February 2018, https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8279452, hereinafter referred to as “Liu”.
Regarding claim 1, Shuqair teaches A system (Shuqair, Page 868, Abstract, Lines 4-7, “This paper proposes a novel architecture based on reinforcement learning (RL) to incrementally learn patterns of time-series data and detect changes in the system state”; see also Shuqair, Page 869, Figure 1; Shuqair, Page 869, Section A, Paragraph 2, Lines 1-4, “We used a one-class SVM classifier to implement each anomaly classification model. One-class classifiers, including one-class SVM, have been used as an unsupervised method for anomaly or outlier detection applications”; “Anomaly Analyzer Unit” is considered to be the “unsupervised model”), comprising:
one or more processors; and one or more non-transitory, computer-readable media comprising instructions that, when executed by the one or more processors, cause operations (Shuqair, Page 873, Section D, Paragraph 2, 11-15, “The coding of the proposed model utilized the use of Python programming language, the implementation of the three classification models and the one class SVMs used the SciKit-Learn library, and the DQN for the RL agent was built with Keras. library”; The use of these tools provides evidence of a generic computer which would comprise “one or more processors”, “one or more non-transitory, computer-readable media”, and “instructions”) comprising:
receiving a first time-series dataset (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
processing, using an unsupervised model that is based on candidate models that are trained independently on separate sets of training data (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2”; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”)…, the first time-series dataset to generate a set of changepoints in the first time-series dataset (Shuqair, Page 868, Section II, Lines 13-16, “A total of 118 features are extracted from every W seconds window of the data stream. The extracted feature vectors are passed to the AAU to produce anomaly scores”; Shuqair, Page 870, Lines 3-6, “They provide a score representing the amount of anomaly in the incoming data stream to the training data. If the anomaly score exceeds a predefined threshold, the algorithm declares a change in the incoming data”);
using the first time-series dataset and the set of changepoints, generating a first training dataset (Shuqair, Page 868, Lines 16-17, “The anomaly scores are then formed as state vectors with six state values in each vector”; Shuqair, Page 869, Figure 1, see the arrow pointing from the “Anomaly Analyzer Unit” to the “Agent”);
using the first training dataset, training a reinforcement learning model, that is distinct from the unsupervised model, to identify changepoints in time-series data (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1, The agent can be seen identifying “transition events” which are considered to be “changepoints”); and
using the reinforcement learning model, processing a second time-series dataset to generate one or more notifications comprising changepoints in the second time-series dataset (Shuqair, Page 869, Paragraph 2, Lines 1-3, “After the training process is complete, the trained RL agent can then be applied to new unseen data to make a classification decision as in Fig 2”).
Shuqair does not explicitly teach the unsupervised model is configured to output a weighted average of a first candidate model of the candidate models and a second candidate model of the candidate models.
Liu teaches the unsupervised model is configured to output a weighted average of a first candidate model of the candidate models and a second candidate model of the candidate models (Liu, Page 576, Col 2, Lines 1-4, “To generate the ensemble models, the outputs of the individual models were then combined using weighted average, i.e., a model with higher performance was assigned a larger weight”).
It would have been obvious, to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the method of Shuqair to include combining the candidate models’ output using a weighted average as taught by Liu. The motivation to do so would have been that weighted average outperforms other ensemble methods (Liu, Page 580, Col 2, Paragraph 2, Lines 21-25, “If the outputs are continuous probabilities, we can use the simple average, weighted average, minimum, maximum, or product to combine the predicted probabilities[68].Studies have indicated that the weighted average approach generally outperforms the simple average approach”).
Regarding claim 2, Shuqair teaches A method for training a reinforcement learning model using training data generated using an unsupervised model (Shuqair, Page 868, Abstract, Lines 4-7, “This paper proposes a novel architecture based on reinforcement learning (RL) to incrementally learn patterns of time-series data and detect changes in the system state”; see also Shuqair, Page 869, Figure 1; Shuqair, Page 869, Section A, Paragraph 2, Lines 1-4, “We used a one-class SVM classifier to implement each anomaly classification model. One-class classifiers, including one-class SVM, have been used as an unsupervised method for anomaly or outlier detection applications”; “Anomaly Analyzer Unit” is considered to be the “unsupervised model”), the method comprising:
processing, using an unsupervised model that is configured to output a combined output of a first model and a second model (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2”; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”; see also Shuqair, Page 871, Figure 3), a first unlabeled dataset to generate a first set of labels associated with statistical properties of the first unlabeled dataset (Shuqair, Page 868, Section II, Lines 13-16, “A total of 118 features are extracted from every W seconds window of the data stream. The extracted feature vectors are passed to the AAU to produce anomaly scores”; Shuqair, Page 870, Lines 3-6, “They provide a score representing the amount of anomaly in the incoming data stream to the training data. If the anomaly score exceeds a predefined threshold, the algorithm declares a change in the incoming data”);
using the first set of labels and the first unlabeled dataset, generating a labeled training dataset (Shuqair, Page 868, Lines 16-17, “The anomaly scores are then formed as state vectors with six state values in each vector”; Shuqair, Page 869, Figure 1, see the arrow pointing from the “Anomaly Analyzer Unit” to the “Agent”);
using the labeled training dataset, training a reinforcement learning model to identify abnormalities and changes to statistical distributions within data (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1, The agent can be seen identifying “transition events” which are considered to be “changepoints”); and
using the reinforcement learning model, processing a second unlabeled dataset to generate a second set of labels associated with statistical properties of the second unlabeled dataset (Shuqair, Page 869, Paragraph 2, Lines 1-3, “After the training process is complete, the trained RL agent can then be applied to new unseen data to make a classification decision as in Fig 2”).
Shuqair does not explicitly teach that the combined output is a weighted average.
Liu teaches a combined output that is a weighted average (Liu, Page 576, Col 2, Lines 1-4, “To generate the ensemble models, the outputs of the individual models were then combined using weighted average, i.e., a model with higher performance was assigned a larger weight”).
It would have been obvious, to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the method of Shuqair to include combining the models’ output using a weighted average as taught by Liu. The motivation to do so would have been that weighted average outperforms other ensemble methods (Liu, Page 580, Col 2, Paragraph 2, Lines 21-25, “If the outputs are continuous probabilities, we can use the simple average, weighted average, minimum, maximum, or product to combine the predicted probabilities[68].Studies have indicated that the weighted average approach generally outperforms the simple average approach”).
Regarding claim 10, the rejection of claim 2 is incorporated, and further, Shuqair teaches wherein the reinforcement learning model performs changepoint detection using a Q-learning algorithm (Shuqair, Page 870, Lines 5-9, “The RL action is the agent’s decision whether the current class label has remained the same. It combines Q-learning [5] with deep artificial neural networks to create a deep Q-network (DQN) model”).
Regarding claim 11, the rejection of claim 2 is incorporated, and further, Shuqair teaches wherein the reinforcement learning model performs anomaly detection using a deep reinforcement learning algorithm (Shuqair, Page 870, Section C, Lines 2-5, “We implemented the RL agent in our approach using deep reinforcement learning, where the RL model employs a deep artificial neural network as its non-linear function estimator to take the optimal actions”).
Regarding claim 12, Shuqair teaches, One or more non-transitory computer-readable media comprising instructions that, when executed by one or more processors, cause operations (Shuqair, Page 873, Section D, Paragraph 2, 11-15, “The coding of the proposed model utilized the use of Python programming language, the implementation of the three classification models and the one class SVMs used the SciKit-Learn library, and the DQN for the RL agent was built with Keras. library”; The use of these tools provides evidence of a generic computer which would comprise “one or more processors”, “one or more non-transitory, computer-readable media”, and “instructions”) comprising:
processing, using a first machine learning model that is configured to output a combined output of a first model and a second model (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2”; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”; see also Shuqair, Page 871, Figure 3), a first unlabeled dataset to generate a first set of labels associated with statistical properties of the first unlabeled dataset (Shuqair, Page 868, Section II, Lines 13-16, “A total of 118 features are extracted from every W seconds window of the data stream. The extracted feature vectors are passed to the AAU to produce anomaly scores”; Shuqair, Page 870, Lines 3-6, “They provide a score representing the amount of anomaly in the incoming data stream to the training data. If the anomaly score exceeds a predefined threshold, the algorithm declares a change in the incoming data”);
using the first set of labels and the first unlabeled dataset, generating a labeled training dataset (Shuqair, Page 868, Lines 16-17, “The anomaly scores are then formed as state vectors with six state values in each vector”; Shuqair, Page 869, Figure 1, see the arrow pointing from the “Anomaly Analyzer Unit” to the “Agent”);
using the labeled training dataset, training a second machine learning model to identify abnormalities and changes to statistical distributions within data (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1, The agent can be seen identifying “transition events” which are considered to be “changepoints”); and
using the second machine learning model, processing a second unlabeled dataset to generate a second set of labels associated with statistical properties of the second unlabeled dataset (Shuqair, Page 869, Paragraph 2, Lines 1-3, “After the training process is complete, the trained RL agent can then be applied to new unseen data to make a classification decision as in Fig 2”).
Shuqair does not explicitly teach that the combined output is a weighted average.
Liu teaches a combined output that is a weighted average (Liu, Page 576, Col 2, Lines 1-4, “To generate the ensemble models, the outputs of the individual models were then combined using weighted average, i.e., a model with higher performance was assigned a larger weight”).
It would have been obvious, to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the method of Shuqair to include combining the models’ output using a weighted average as taught by Liu. The motivation to do so would have been that weighted average outperforms other ensemble methods (Liu, Page 580, Col 2, Paragraph 2, Lines 21-25, “If the outputs are continuous probabilities, we can use the simple average, weighted average, minimum, maximum, or product to combine the predicted probabilities[68].Studies have indicated that the weighted average approach generally outperforms the simple average approach”).
Regarding claim 20, the rejection of claim 12 is incorporated, and further, Shuqair teaches wherein the second machine learning model performs changepoint detection using a Q-learning algorithm (Shuqair, Page 870, Lines 5-9, “The RL action is the agent’s decision whether the current class label has remained the same. It combines Q-learning [5] with deep artificial neural networks to create a deep Q-network (DQN) model”).
Claims 3, 9, 13, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Shuqair in view of Liu in further view of Gedda et al., Interactive Change Point Detection using optimisation approach and Bayesian statistics applied to real world applications, 06/17/2021, https://arxiv.org/pdf/2106.09691, hereinafter referred to as “Gedda”.
Regarding claim 3, the rejection of claim 2 is incorporated, the proposed combination teaches the first unlabeled dataset comprises time-series data (Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”).
The proposed combination does not explicitly teach the unsupervised model uses a Bayesian network to perform changepoint detection on time-series data and generate timestamps corresponding to points in time-series data where a statistical distribution has shifted.
Gedda teaches the unsupervised model uses a Bayesian network to perform changepoint detection on time-series data and generate timestamps corresponding to points in time-series data where a statistical distribution has shifted (Gedda, Page 10, Section 2.3, Lines 1-5, “In contrast to the optimisation approach, the Bayesian approach is based on Bayes’ probability theorem, where the maximum probabilities are identified. It is based on the Bayesian principle of calculating a posterior distribution of a time stamp being a change point, given a prior and a likelihood function. From this posterior distribution, we can identify the points which are most likely to be change points”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the unsupervised model of the proposed combination to include a Bayesian network to perform changepoint detection. The motivation to do so would have been that a Bayesian approach uses prior knowledge and achieves high accuracy on a wide variety of datasets (Gedda, Pages 22-33, Results; Gedda, Page 43, Section 6, Lines 7-9, “The Bayesian approach uses prior knowledge on the distance between change points and a likelihood function with information about the segments in order to predict the probability of a time point being a change point”).
Regarding claim 9, the rejection of claim 2 is incorporated, and further, the proposed combination teaches updating the reinforcement learning model based on the … training dataset (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1).
The proposed combination does not explicitly teach presenting the second set of labels to a set of users; obtaining a set of feedback from the set of users, wherein the set of feedback is indicative of a degree of suitability of the second set of labels to the second unlabeled dataset; using the set of feedback, generate a second training dataset, wherein the second training dataset is labeled using the set of feedback.
Gedda teaches presenting the second set of labels to a set of users; obtaining a set of feedback from the set of users, wherein the set of feedback is indicative of a degree of suitability of the second set of labels to the second unlabeled dataset; using the set of feedback, generate a second training dataset, wherein the second training dataset is labeled using the set of feedback (Gedda, Page 41, Section 5.4, Lines 1-3, “we want to explore the possibilities of incorporating user feedback or prior knowledge. This can either be done via changes in settings of the unsupervised approaches or after predictions have been made”; Gedda, Page 42, Section 5.4.2, Paragraph 3, Lines 4-8, “This means that if a user knows approximately where a change point is present, a probability distribution with mean and variance according to the users’ expectation can be found and joined with the calculated posterior distribution. This means that the posterior distribution of change points can incorporate the information from users without being recalculated”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the training dataset of the proposed combination to include modifying the dataset using user feedback as taught by Gedda. The motivation to do so would have been to make the approach more user friendly and increase the method’s effectiveness (Gedda, Page 41, Section 5.4, Lines 1-2, “To make the explored unsupervised CPD approaches user friendly and more effective, we want to explore the possibilities of incorporating user feedback or prior knowledge”).
Regarding claim 13, the rejection of claim 12 is incorporated, the proposed combination teaches the first unlabeled dataset comprises time-series data (Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”).
The proposed combination does not explicitly teach the first machine learning model uses a Bayesian network to perform changepoint detection on time-series data and generate timestamps corresponding to points in time-series data where a statistical distribution shifted.
Gedda teaches the first machine learning model uses a Bayesian network to perform changepoint detection on time-series data and generate timestamps corresponding to points in time-series data where a statistical distribution shifted (Gedda, Page 10, Section 2.3, Lines 1-5, “In contrast to the optimisation approach, the Bayesian approach is based on Bayes’ probability theorem, where the maximum probabilities are identified. It is based on the Bayesian principle of calculating a posterior distribution of a time stamp being a change point, given a prior and a likelihood function. From this posterior distribution, we can identify the points which are most likely to be change points”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the unsupervised model of the proposed combination to include a Bayesian network to perform changepoint detection. The motivation to do so would have been that a Bayesian approach uses prior knowledge and achieves high accuracy on a wide variety of datasets (Gedda, Pages 22-33, Results; Gedda, Page 43, Section 6, Lines 7-9, “The Bayesian approach uses prior knowledge on the distance between change points and a likelihood function with information about the segments in order to predict the probability of a time point being a change point”).
Regarding claim 19, the rejection of claim 12 is incorporated, and further, the proposed combination teaches updating the second machine learning model based on the … training dataset (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1).
The proposed combination does not explicitly teach presenting the second set of labels to a set of users; collecting a set of feedback from the set of users, wherein the set of feedback is indicative of a degree of suitability of the second set of labels to the second unlabeled dataset; using the set of feedback, generate a second training dataset, wherein the second training dataset is labeled using the set of feedback.
Gedda teaches presenting the second set of labels to a set of users; collecting a set of feedback from the set of users, wherein the set of feedback is indicative of a degree of suitability of the second set of labels to the second unlabeled dataset; using the set of feedback, generate a second training dataset, wherein the second training dataset is labeled using the set of feedback (Gedda, Page 41, Section 5.4, Lines 1-3, “we want to explore the possibilities of incorporating user feedback or prior knowledge. This can either be done via changes in settings of the unsupervised approaches or after predictions have been made”; Gedda, Page 42, Section 5.4.2, Paragraph 3, Lines 4-8, “This means that if a user knows approximately where a change point is present, a probability distribution with mean and variance according to the users’ expectation can be found and joined with the calculated posterior distribution. This means that the posterior distribution of change points can incorporate the information from users without being recalculated”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the training dataset of the proposed combination to include modifying the dataset using user feedback as taught by Gedda. The motivation to do so would have been to make the approach more user friendly and increase the method’s effectiveness (Gedda, Page 41, Section 5.4, Lines 1-2, “To make the explored unsupervised CPD approaches user friendly and more effective, we want to explore the possibilities of incorporating user feedback or prior knowledge”).
Claims 4-7 and 14-17 are rejected under 35 U.S.C. 103 as being unpatentable over Shuqair in view of Liu in further view of Yu-Quan et al., Dynamic weighting ensemble classifiers based on cross-validation, 05/04/2010, https://www.researchgate.net/publication/220372354_Dynamic_weighting_ensemble_classifiers_based_on_cross-validation, hereinafter referred to as “Yu-Quan”.
Regarding claim 4, the rejection of claim 2 is incorporated, and further, the proposed combination teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
partitioning the raw training data into a first portion … and a second portion … (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
training the first model using the first portion of the raw training data to obtain a first set of model parameters; training the second model using the second portion of the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”); and
generating the unsupervised model by combining the first set of model parameters and the second set of model parameters … (Shuqair, Page 869, Section A, Lines 20-25, “The three reported anomaly scores will be generated continuously every step T seconds. The scores are then concatenated into one time-series anomaly score stream to form the output of AUU. Fig. 3 shows the overall structure of AAU and a sample anomaly score stream”).
The proposed combination does not explicitly teach associating the training data with weights nor combining the parameters based on the first weight score and the second weight score.
Yu-Quan teaches associating the training data with weights nor combining the parameters based on the first weight score and the second weight score (Yu-Quan, Page 311, Section 3, Lines 2-5, “(1) utilize Random Subspace approach to train individual classifiers; (2) employ cross-validation technique to dynamically assign a weight to each base classifier and (3) combine the output of each classifier with the corresponding weight to give the final prediction”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include weighting the training data and combining the parameters of the models based on the weights as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Regarding claim 5, the rejection of claim 2 is incorporated, and further, Shuqair teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
selecting the first model for unsupervised learning and the second model for unsupervised learning (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
training the first model using the raw training data to obtain a first set of model parameters; training the second model using the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”); and
generating the unsupervised model by combining the first set of model parameters and the second set of model parameters… (Shuqair, Page 869, Section A, Lines 20-25, “The three reported anomaly scores will be generated continuously every step T seconds. The scores are then concatenated into one time-series anomaly score stream to form the output of AUU. Fig. 3 shows the overall structure of AAU and a sample anomaly score stream”).
The proposed combination does not explicitly teach combining the model parameters based on a first bias metric.
Yu-Quan teaches combining the model parameters based on a first bias metric (Yu-Quan, Page 311, Section 3, Lines 2-5, “(1) utilize Random Subspace approach to train individual classifiers; (2) employ cross-validation technique to dynamically assign a weight to each base classifier and (3) combine the output of each classifier with the corresponding weight to give the final prediction” The “weight” assigned is considered to be the bias metric as the overall architecture is “biased” toward models with lower cross-validation error).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include combining the parameters of the models based on the bias metric as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Regarding claim 6, the rejection of claim 5 is incorporated, and further, the proposed combination teaches based on a performance metric of the reinforcement learning model, updating the unsupervised model (Shuqair, Page 868, Section II, Lines 21-23, “The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken”; Shuqair, Page 869, Paragraph 2, Lines 6-8, “The parameters of the AAU are adjusted following the RL agent’s actions”).
Yu-Quan teaches updating the first bias metric to generate a second bias metric (Yu-Quan, Page 312, Section 3.2, Lines 16-25, “construct N classifiers in N feature subspaces accordingly (these classifiers are trained using training set excluding validation samples in each validating process) and utilize these classifier to classify samples in the valid neighborhood. If the classifier in a certain feature subspace can correctly classify all these neighbors, then the value of the weight of the corresponding subspace pluses 1. When above mentioned procedure is performed M times, the weight of each subspace is finally obtained”; When the parameters of the AAU are updated, cross validation would need to be performed again, thus changing the “weight of each subspace”); and
updating the unsupervised model by combining the first set of model parameters and the second set of model parameters based on the second bias metric (Yu-Quan, Page 312, Section 3.2, Paragraph 2, “Finally, all the binary function outputs and weights of individual classifiers in different feature subspaces are combined using the rule of majority voting to give the final predicted label as follows
Ω
e
n
s
X
t
e
s
t
=
a
r
g
max
ω
l
∑
j
=
1
N
W
j
(
X
t
e
s
t
)
θ
j
(
X
t
e
s
t
∈
ω
l
)
”).
Regarding claim 7, the rejection of claim 2 is incorporated, and further, Shuqair teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
selecting the first model for unsupervised learning and the second model for unsupervised learning (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class” The class 1 and 2 anomaly detectors are considered to be the “first model for unsupervised learning and a second model for unsupervised learning”);
training the first model using the raw training data to obtain a first set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
…
training the second model using the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”).
The proposed combination does not explicitly teach determining a first cross-validation error score associated with the first set of model parameters, wherein the first cross-validation error score is indicative of a degree of fit between the first model and the raw training data; determining a second cross-validation error score associated with the second set of model parameters, wherein the second cross-validation error score is indicative of a degree of fit between the second model and the raw training data; and generating the unsupervised model by combining the first set of model parameters and the second set of model parameters based on the first cross-validation error score and the second cross-validation error score.
Yu-Quan teaches determining a first cross-validation error score associated with the first set of model parameters, wherein the first cross-validation error score is indicative of a degree of fit between the first model and the raw training data; determining a second cross-validation error score associated with the second set of model parameters, wherein the second cross-validation error score is indicative of a degree of fit between the second model and the raw training data (Yu-Quan, Page 312, Section 3.2, Lines 1-25, “Now, let the class label set is {x1, x2, …, xL}. N feature subspaces are produced by Random Subspace method, and corresponding weight set of these subspaces is W ={W1, W2, …, WN}. The weight of each subspace is initialized to 0. After that, equally partition the training data TrainSum into M different folds which are stratified so that the class distribution of the samples in each fold is approximately the same as that in the initial data. M-1 out of M of the folds are used to train the classifier, and the remaining fold is used as the validation set. The procedure is repeated M times, with a different fold being used as the validation set. In validating iteration i (
1
≤
i
≤
M
), for a new test sample Xtest, first seek K (K is the size of neighborhood and set up manually) nearest neighbors from the validation set, then wipe off false neighbors to form the valid neighborhood using (3). Next, construct N classifiers in N feature subspaces accordingly (these classifiers are trained using training set excluding validation samples in each validating process) and utilize these classifier to classify samples in the valid neighborhood. If the classifier in a certain feature subspace can correctly classify all these neighbors, then the value of the weight of the corresponding subspace pluses 1. When above mentioned procedure is performed M times, the weight of each subspace is finally obtained”); and
generating the unsupervised model by combining the first set of model parameters and the second set of model parameters based on the first cross-validation error score and the second cross-validation error score (Yu-Quan, Page 312, Section 3.2, Paragraph 2, “Finally, all the binary function outputs and weights of individual classifiers in different feature subspaces are combined using the rule of majority voting to give the final predicted label as follows
Ω
e
n
s
X
t
e
s
t
=
a
r
g
max
ω
l
∑
j
=
1
N
W
j
(
X
t
e
s
t
)
θ
j
(
X
t
e
s
t
∈
ω
l
)
”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include combining the parameters of the models based on the cross-validation error scores as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Regarding claim 14, the rejection of claim 12 is incorporated, and further, Shuqair teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
partitioning the raw training data into a first portion … and a second portion … (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
training the first model using the first portion of the raw training data to obtain a first set of model parameters; training the second model using the second portion of the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”); and
generating the first machine learning model by combining the first set of model parameters and the second set of model parameters … (Shuqair, Page 869, Section A, Lines 20-25, “The three reported anomaly scores will be generated continuously every step T seconds. The scores are then concatenated into one time-series anomaly score stream to form the output of AUU. Fig. 3 shows the overall structure of AAU and a sample anomaly score stream”).
The proposed combination does not explicitly teach associating the training data with weights nor combining the parameters based on the first weight score and the second weight score.
Yu-Quan teaches associating the training data with weights nor combining the parameters based on the first weight score and the second weight score (Yu-Quan, Page 311, Section 3, Lines 2-5, “(1) utilize Random Subspace approach to train individual classifiers; (2) employ cross-validation technique to dynamically assign a weight to each base classifier and (3) combine the output of each classifier with the corresponding weight to give the final prediction”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include weighting the training data and combining the parameters of the models based on the weights as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Regarding claim 15, the rejection of claim 12 is incorporated, and further, Shuqair teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
selecting the first model for unsupervised learning and the second model for unsupervised learning (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
training the first model using the raw training data to obtain a first set of model parameters; training a second model using the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”); and
generating the first machine learning model by combining the first set of model parameters and the second set of model parameters… (Shuqair, Page 869, Section A, Lines 20-25, “The three reported anomaly scores will be generated continuously every step T seconds. The scores are then concatenated into one time-series anomaly score stream to form the output of AUU. Fig. 3 shows the overall structure of AAU and a sample anomaly score stream”).
The proposed combination does not explicitly teach combining the model parameters based on a first bias metric.
Yu-Quan teaches combining the model parameters based on a first bias metric (Yu-Quan, Page 311, Section 3, Lines 2-5, “(1) utilize Random Subspace approach to train individual classifiers; (2) employ cross-validation technique to dynamically assign a weight to each base classifier and (3) combine the output of each classifier with the corresponding weight to give the final prediction” The “weight” assigned is considered to be the bias metric as the overall architecture is “biased” toward models with lower cross-validation error).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include combining the parameters of the models based on the bias metric as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Regarding claim 16, the rejection of claim 15 is incorporated, and further, the proposed combination teaches based on a performance metric of the second machine learning model, updating the unsupervised model (Shuqair, Page 868, Section II, Lines 21-23, “The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken”; Shuqair, Page 869, Paragraph 2, Lines 6-8, “The parameters of the AAU are adjusted following the RL agent’s actions”).
Yu-Quan teaches updating the first bias metric to generate a second bias metric (Yu-Quan, Page 312, Section 3.2, Lines 16-25, “construct N classifiers in N feature subspaces accordingly (these classifiers are trained using training set excluding validation samples in each validating process) and utilize these classifier to classify samples in the valid neighborhood. If the classifier in a certain feature subspace can correctly classify all these neighbors, then the value of the weight of the corresponding subspace pluses 1. When above mentioned procedure is performed M times, the weight of each subspace is finally obtained”; When the parameters of the AAU are updated, cross validation would need to be performed again, thus changing the “weight of each subspace”); and
updating the first machine learning model by combining the first set of model parameters and the second set of model parameters based on the second bias metric (Yu-Quan, Page 312, Section 3.2, Paragraph 2, “Finally, all the binary function outputs and weights of individual classifiers in different feature subspaces are combined using the rule of majority voting to give the final predicted label as follows
Ω
e
n
s
X
t
e
s
t
=
a
r
g
max
ω
l
∑
j
=
1
N
W
j
(
X
t
e
s
t
)
θ
j
(
X
t
e
s
t
∈
ω
l
)
”).
Regarding claim 17, the rejection of claim 12 is incorporated, and further, Shuqair teaches receiving raw training data (Shuqair, Page 868, Section II, Lines 12-13, “The incoming data is fed to the model as a data stream during the training process, Fig. 1”; Shuqair, Page 868, Section I, Paragraph 3, Lines 4-6, “We evaluated the generalizability of the proposed classification model on time series wearable data from a Parkinson’s disease (PD) dataset”);
selecting the first model for unsupervised learning and the second model for unsupervised learning (Shuqair, Page 869, Section A, Lines 2-8, “For this purpose, we constructed the AAU using three anomaly detection classifiers where each classifier extracts changes in different aspects of the data stream. One is to represent how the overall data stream changes over time. The other two also characterize changes over time, but one with relation to class 1 and one for class 2; Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class” The class 1 and 2 anomaly detectors are considered to be the “first model for unsupervised learning and a second model for unsupervised learning”);
training the first model using the raw training data to obtain a first set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”);
…
training the second model using the raw training data to obtain a second set of model parameters (Shuqair, Page 869, Section A, Lines 17-20, “On the other hand, class 1 and 2 anomaly detectors will be trained and updated using only their class data, so their reported anomaly scores will describe the data dynamics of each class”).
The proposed combination does not explicitly teach determining a first cross-validation error score associated with the first set of model parameters, wherein the first cross-validation error score is indicative of a degree of fit between the first model and the raw training data; determining a second cross-validation error score associated with the second set of model parameters, wherein the second cross-validation error score is indicative of a degree of fit between the second model and the raw training data; and generating the first machine learning model by combining the first set of model parameters and the second set of model parameters based on the first cross-validation error score and the second cross-validation error score.
Yu-Quan teaches determining a first cross-validation error score associated with the first set of model parameters, wherein the first cross-validation error score is indicative of a degree of fit between the first model and the raw training data; determining a second cross-validation error score associated with the second set of model parameters, wherein the second cross-validation error score is indicative of a degree of fit between the second model and the raw training data (Yu-Quan, Page 312, Section 3.2, Lines 1-25, “Now, let the class label set is {x1, x2, …, xL}. N feature subspaces are produced by Random Subspace method, and corresponding weight set of these subspaces is W ={W1, W2, …, WN}. The weight of each subspace is initialized to 0. After that, equally partition the training data TrainSum into M different folds which are stratified so that the class distribution of the samples in each fold is approximately the same as that in the initial data. M-1 out of M of the folds are used to train the classifier, and the remaining fold is used as the validation set. The procedure is repeated M times, with a different fold being used as the validation set. In validating iteration i (
1
≤
i
≤
M
), for a new test sample Xtest, first seek K (K is the size of neighborhood and set up manually) nearest neighbors from the validation set, then wipe off false neighbors to form the valid neighborhood using (3). Next, construct N classifiers in N feature subspaces accordingly (these classifiers are trained using training set excluding validation samples in each validating process) and utilize these classifier to classify samples in the valid neighborhood. If the classifier in a certain feature subspace can correctly classify all these neighbors, then the value of the weight of the corresponding subspace pluses 1. When above mentioned procedure is performed M times, the weight of each subspace is finally obtained”); and
generating the first machine learning model by combining the first set of model parameters and the second set of model parameters based on the first cross-validation error score and the second cross-validation error score (Yu-Quan, Page 312, Section 3.2, Paragraph 2, “Finally, all the binary function outputs and weights of individual classifiers in different feature subspaces are combined using the rule of majority voting to give the final predicted label as follows
Ω
e
n
s
X
t
e
s
t
=
a
r
g
max
ω
l
∑
j
=
1
N
W
j
(
X
t
e
s
t
)
θ
j
(
X
t
e
s
t
∈
ω
l
)
”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to have modified the generation of the unsupervised model as taught by the proposed combination, to include combining the parameters of the models based on the cross-validation error scores as taught by Yu-Quan. The motivation to do so would have been that the introduction of cross-validation improves the integrated model performance (Yu-Quan, Page 311, Col 2, Lines 19-24, “The introduction of cross-validation not only takes into consideration the distinguished abilities of classifiers in different feature subspaces, but also makes the validation set can be reused to train base classifiers, thus does not affect the precision of the training model and improves the integrated performance”).
Claims 8 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Shuqair in view of Liu in further view of Semenoglou et al., Data augmentation for univariate time series forecasting with neural networks, 10/28/2022, https://doi.org/10.1016/j.patcog.2022.109132, hereinafter referred to as “Semenoglou”.
Regarding claim 8, the rejection of claim 2 is incorporated, and further, Shuqair teaches using the … set of labels and the … unlabeled dataset, generate an expanded training dataset (Shuqair, Page 868, Lines 16-17, “The anomaly scores are then formed as state vectors with six state values in each vector”; Shuqair, Page 869, Figure 1, see the arrow pointing from the “Anomaly Analyzer Unit” to the “Agent”); and
using the … training dataset, training the reinforcement learning model (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1).
The proposed combination does not explicitly teach combining the first unlabeled dataset with synthetic data to generate an augmented unlabeled dataset, wherein the synthetic data is representative of hypothetical scenarios not included in the first unlabeled dataset; processing the augmented unlabeled dataset using the unsupervised model to generate an expanded set of labels.
Semenoglou teaches combining the first unlabeled dataset with synthetic data to generate an augmented unlabeled dataset, wherein the synthetic data is representative of hypothetical scenarios not included in the first unlabeled dataset (Semenoglou, Page 5, Section 3.8, Lines 1-6, “In addition to the existing data augmentation techniques de- scribed earlier, we introduce a novel approach that re-samples the original samples of the train set to create new ones that closely resemble the originals but focus on particular parts of them with the objective to emphasize local variations of the series that may have otherwise been overlooked”; Semenoglou, Page 7, Section 4.4, Lines 1-2, “For each set, we create synthetic samples so that the initial number of training samples doubles”);
processing the augmented unlabeled dataset using the unsupervised model to generate an expanded set of labels (Semenoglou, Page 7, Section 4.4, Paragraph 2, Lines 9-11, “The resulting set, containing the synthetic samples along with the originals, is the final train set used by the NN models”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention to include modifying the unlabeled dataset of the proposed combination to include augmenting synthetic data as taught by Semenoglou. The motivation to do so would have been to increase the size of the training dataset, improving the accuracy of the resulting model (Semenoglou, Page 1, Abstract, Lines 1-4, “Neural networks have been proven particularly accurate in univariate time series forecasting settings, re- quiring however a significant number of training samples to be effectively trained. In machine learning applications where available data are limited, data augmentation techniques have been successfully used to generate synthetic data that resemble and complement the original train set”).
Regarding claim 18, the rejection of claim 12 is incorporated, and further, Shuqair teaches using the … set of labels and the … unlabeled dataset, generate an expanded training dataset (Shuqair, Page 868, Lines 16-17, “The anomaly scores are then formed as state vectors with six state values in each vector”; Shuqair, Page 869, Figure 1, see the arrow pointing from the “Anomaly Analyzer Unit” to the “Agent”); and
using the … training dataset, training the second machine learning model (Shuqair, Page 868, Section II, Line 17 – Page 869, Line 4, “The RL agent observes these scores over time and decides whether the incoming data belongs to the same class label or if there has been a substantial change in the data dynamics to signal a transition to the other class label. The RL agent changes the environment states by adjusting the AAU parameters according to the actions taken. These transitions (i.e., RL agent actions) are then interpreted as the model predicted class labels for the incoming data stream. The environment rewards the RL agent during the training process corresponding to the actions made. This process is repeated until the RL agent learns to make the optimal decision about the classification labels”; see also Shuqair, Page 869, Figure 1).
The proposed combination does not explicitly teach combining the first unlabeled dataset with synthetic data to generate an augmented unlabeled dataset, wherein the synthetic data is representative of hypothetical scenarios not included in the first unlabeled dataset; processing the augmented unlabeled dataset using the first machine learning model to generate an expanded set of labels.
Semenoglou teaches combining the first unlabeled dataset with synthetic data to generate an augmented unlabeled dataset, wherein the synthetic data is representative of hypothetical scenarios not included in the first unlabeled dataset (Semenoglou, Page 5, Section 3.8, Lines 1-6, “In addition to the existing data augmentation techniques de- scribed earlier, we introduce a novel approach that re-samples the original samples of the train set to create new ones that closely resemble the originals but focus on particular parts of them with the objective to emphasize local variations of the series that may have otherwise been overlooked”; Semenoglou, Page 7, Section 4.4, Lines 1-2, “For each set, we create synthetic samples so that the initial number of training samples doubles”);
processing the augmented unlabeled dataset using the first machine learning model to generate an expanded set of labels (Semenoglou, Page 7, Section 4.4, Paragraph 2, Lines 9-11, “The resulting set, containing the synthetic samples along with the originals, is the final train set used by the NN models”).
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention to include modifying the unlabeled dataset of the proposed combination to include augmenting synthetic data as taught by Semenoglou. The motivation to do so would have been to increase the size of the training dataset, improving the accuracy of the resulting model (Semenoglou, Page 1, Abstract, Lines 1-4, “Neural networks have been proven particularly accurate in univariate time series forecasting settings, re- quiring however a significant number of training samples to be effectively trained. In machine learning applications where available data are limited, data augmentation techniques have been successfully used to generate synthetic data that resemble and complement the original train set”).
Response to Arguments
Applicant’s arguments regarding the prior art rejections of the claims have been fully considered but are unpersuasive.
Applicant’s arguments with respect to the amendments to the claims, specifically the weighted combination of models, 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.
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.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MOLLY CLARKE SIPPEL whose telephone number is (571)272-3270. The examiner can normally be reached Monday - Friday, 7:30 a.m. - 4:30 p.m. ET..
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, Kakali Chaki can be reached at (571)272-3719. 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.
/M.C.S./ Examiner, Art Unit 2122
/KAKALI CHAKI/ Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights