DETAILED ACTION
Notice of 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 .
Information Disclosure Statement
The information disclosure statements submitted on 5/3/2023, 10/17/2024, and 11/18/2025 have been considered.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f):
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f). The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f) is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f). The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 U.S.C. 112(f) except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) except as otherwise indicated in an Office action.
This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) because the claim limitation(s) uses a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier. Such claim limitations (and associated disclosure) is/are:
Limitation
Applicable Claims
Disclosure
a data extractor ... configured to extract test data and training data from an ordered data stream
8-14
Paras. 0004, 0062
a doubly robust causal learning outcome predictor ... configured to predict doubly robust outcomes for the ordered data stream ...
8-14
Para. 0010, 0043-0044
a concept drift detector ... configured to measure concept drift between the test data and the training data ...
8-14
Fig. 1, Concept Drive Detector 110; Fig. 2, Concept Drift Detector 208, paras. 0027-0028, 0032-0038
an adversarial feature selector ... configured to select retraining feature vectors ...
8-14
Fig. 2, Adversarial Feature Selector 212, paras. 0039-0041
a machine learning model retrainer ... configured to retrain the machine learning model using the retraining feature vectors
8-14
Fig. 1, Machine Learning Model Trainer 112; Fig. 2, Machine Learning Model Trainer 214, paras. 0033, 0042
Because these claim limitations are being interpreted under 35 U.S.C. 112(f), they are being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof.
If applicant does not intend to have these limitations interpreted under 35 U.S.C. 112(f), applicant may: (1) amend the claim limitations to avoid them being interpreted under 35 U.S.C. 112(f) (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitations recite sufficient structure to perform the claimed function so as to avoid them being interpreted under 35 U.S.C. 112(f).
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 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding Step 1 of the Alice/Mayo framework, Claims 1-7 are directed to a method (a process), Claims 8-14 are directed to a system (a machine), and Claims 15-20 are directed to one or more tangible processor-readable storage media (an article of manufacture), which each fall within one of the four statutory categories of inventions.
Regarding Claim 1
Step 2A, prong 1 (Is the claim directed to a law of nature, a natural phenomenon or an abstract idea).
Claim 1 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including an observation, evaluation, judgment, opinion) or with the aid of pencil and paper but for the recitation of generic computer components (e.g., “ordered data stream”, “machine learning model”).
A method of managing model drift in a machine learning model, the method comprising: (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can manage model drift in machine learning models, for example, such a machine learning engineer could determine if drift is cased by either data or concept drift, and then choose to retrain the machine learning model only using training data available after a specific event occurred)
predicting doubly robust outcomes for the ordered data stream based on a combination of treatments predicted based on controls of the ordered data stream and outcomes predicted based on the treatments and the controls of the ordered data stream, wherein the doubly robust outcomes include doubly robust outcomes for the test data and doubly robust outcomes for the training data; (under the broadest reasonable interpretation, a human, such as physician or medical researcher, can predict doubly robust outcomes for data (such as patient data from a patient monitor) to predict treatment options based on control group information)
measuring concept drift between the test data and the training data with respect to the machine learning model, wherein the concept drift is measured as an expectation of differences between the doubly robust outcomes for the ordered data stream and the doubly robust outcomes for the training data; (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can measure concept drift using pencil and paper, using the mathematical equations described in paras. 0032-0033 and 0037-0039 of the instant specification)
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can select certain feature vectors from a group of feature vectors created from the ordered stream data, when the machine learning engineer decides that a concept drive measurement exceeds a threshold for retraining)
Step 2A, prong 2 (Does the claim recite additional elements that integrate the judicial exception into a practical application?).
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional elements (e.g., “ordered data stream”, “machine learning model”) which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “extracting test data and training data from an ordered data stream, the test data being extracted from a detection window in the ordered data stream and the training data being extracted from a sliding reference window that precedes the detection window in the ordered data stream” limitation, such additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Regarding the “retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of retraining a machine learning model. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (any generic training of a machine learning model). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, after considering all claim elements individually and as an ordered combination, it is determined that the claims do not integrate the judicial exception into a practical application.
Step 2B (Does the claim recite additional elements that amount to significantly more than the judicial exception?)
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional elements (e.g., “ordered data stream”, “machine learning model”) are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “extracting test data and training data from an ordered data stream, the test data being extracted from a detection window in the ordered data stream and the training data being extracted from a sliding reference window that precedes the detection window in the ordered data stream” limitation, as discussed above, the additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Regarding the “retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Accordingly, at Step 2B, after considering all claim elements individually and as an ordered combination, it is determined that the claims do not amount to significantly more than the judicial exception.
Regarding Claim 2
Step 2A, Prong 1
the doubly robust outcomes predicted for the ordered data stream are based on the outcomes predicted based on the treatments and the controls of the ordered data stream added to an inverse propensity weighting of residuals between the observed outcomes and the outcomes predicted based on the treatments and the controls of the ordered data stream (under the broadest reasonable interpretation, a human, such as physician or medical researcher, can predict doubly robust outcomes for data (such as patient data from a patient monitor) to predict outcomes (e.g., recovery, death) based on the treatments and control group information, and can further perform an inverse propensity weighting of residuals on paper using the equations in para. 0037 of the instant specification)
Step 2A, Prong 2
Regarding the “wherein the training data includes observed outcomes” limitation, such limitation merely describes the types of training data that will be used to retrain a machine learning model, and therefore such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use. As explained by the Supreme Court, a claim directed to a judicial exception cannot be made eligible "simply by having the applicant acquiesce to limiting the reach of the patent for the formula to a particular technological use." Diamond v. Diehr, 450 U.S. 175, 192 n.14, 209 USPQ 1, 10 n. 14 (1981). Thus, limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not integrate a judicial exception into a practical application.
Step 2B
Regarding the “wherein the training data includes observed outcomes” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use as explained above, which does not amount to significantly more than the judicial exception. MPEP 2106.05(h).
Regarding Claim 3
Step 2A, Prong 1
wherein the inverse propensity weighting of the residuals is based on the treatments predicted based on the controls of the ordered data stream. (under the broadest reasonable interpretation, a human, such as physician or medical researcher, can perform an inverse propensity weighting of residuals on paper using the equations in para. 0037 of the instant specification)
Regarding Step 2A, Prong 2, the claim does not include any additional elements that integrate the judicial exception into a practical application and regarding Step 2B, there are no additional elements recited that amount to significantly more than the judicial exception.
Regarding Claim 4
Step 2A, Prong 1
wherein the selecting operation comprises: selecting the retraining feature vectors based on scores generated by an adversarial feature classifier. (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can select certain feature vectors from a group of feature vectors created from the ordered stream data, where such selection takes into consideration scores generated by an adversarial feature classifier)
Regarding Step 2A, Prong 2, the claim does not include any additional elements that integrate the judicial exception into a practical application and regarding Step 2B, there are no additional elements recited that amount to significantly more than the judicial exception.
Regarding Claim 5
Step 2A, Prong 1
selecting the retraining feature vectors based on an area-under-the-curve (AUC) score generated by the adversarial feature classifier and an AUC condition hyperparameter. (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can select certain feature vectors from a group of feature vectors created from the ordered stream data, where such selection takes into consideration AUC scores and hyperparameters, which are merely values for consideration)
Regarding Step 2A, Prong 2, the claim does not include any additional elements that integrate the judicial exception into a practical application and regarding Step 2B, there are no additional elements recited that amount to significantly more than the judicial exception.
Regarding Claim 6
Step 2A, Prong 1
wherein the selecting operation comprises: selecting the retraining feature vectors based on a raw feature importance value and a raw feature importance condition hyperparameter. (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can select certain feature vectors from a group of feature vectors created from the ordered stream data, where such selection takes into consideration a raw feature importance value and related hyperparameter, which are merely values for consideration)
Regarding Step 2A, Prong 2, the claim does not include any additional elements that integrate the judicial exception into a practical application and regarding Step 2B, there are no additional elements recited that amount to significantly more than the judicial exception.
Regarding Claim 7
Step 2A, Prong 1
wherein the selecting operation comprises: selecting the retraining feature vectors based on a raw permutation feature importance value and a raw permutation feature importance condition hyperparameter. (under the broadest reasonable interpretation, a human, such as an machine learning engineer, can select certain feature vectors from a group of feature vectors created from the ordered stream data, where such selection takes into consideration a raw permutation feature importance value and related hyperparameter, which are merely values for consideration)
Regarding Step 2A, Prong 2, the claim does not include any additional elements that integrate the judicial exception into a practical application and regarding Step 2B, there are no additional elements recited that amount to significantly more than the judicial exception.
Regarding Claim 8
Step 2A, Prong 1
Claim 8 recites a computing system that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 1 with respect to claim 1 also applies to this claim 8. While claim 8 recites additional generic computing components (“hardware processor”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2A, Prong 1.
Step 2A, Prong 2
Claim 8 recites a computing system that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 2 with respect to claim 1 also applies to this claim 8. While claim 8 recites additional generic computing components (“hardware processor”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2A, Prong 2. Such limitations are recited at a high-level of generality and amount to no more than adding the words “apply it” (or an equivalent) with the judicial exception. These additional elements are recited at a high-level of generality and amount to no more than mere instructions to apply the exception using generic computer components. Accordingly, these additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Step 2B
Claim 8 recites a computing system that corresponds to the method of claim 1, and therefore the analysis under Step 2B with respect to claim 1 also applies to this claim 8. While claim 8 recites additional generic computing components (“hardware processor”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2B. Such limitation are recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitations merely provide instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Claims 9-14 depend from claim 8 and correspond to the methods of claims 2-7, respectively, and are therefore rejected for the same reasons explained above with respect to claim 8 and claims 2-7, respectively.
Regarding Claim 15
Step 2A, Prong 1
Claim 15 recites one or more tangible processor-readable storage media that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 1 with respect to claim 1 also applies to this claim 15. While claim 15 recites additional generic computing components (“processors and circuits”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2A, Prong 1.
Step 2A, Prong 2
Claim 15 recites one or more tangible processor-readable storage media that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 2 with respect to claim 1 also applies to this claim 15. While claim 15 recites additional generic computing components (“processors and circuits”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2A, Prong 2. Such limitations are recited at a high-level of generality and amount to no more than adding the words “apply it” (or an equivalent) with the judicial exception. These additional elements are recited at a high-level of generality and amount to no more than mere instructions to apply the exception using generic computer components. Accordingly, these additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Step 2B
Claim 15 recites one or more tangible processor-readable storage media that corresponds to the method of claim 1, and therefore the analysis under Step 2B with respect to claim 1 also applies to this claim 15. While claim 15 recites additional generic computing components (“processors and circuits”, “ordered data stream”, “machine learning model”), such additional generic computing components do not change the analysis under Step 2B. Such limitation are recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitations merely provide instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Claims 16-20 depend from claim 15 and correspond to the methods of claims 2-5 and 7, respectively, and are therefore rejected for the same reasons explained above with respect to claim 15 and claims 2-5 and 7, respectively.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 4-6, 8, 11-13, 15, and 18-19 are rejected under 35 U.S.C. 103 as being unpatentable over Pan, Jing, et al. "Adversarial validation approach to concept drift problem in user targeting automation systems at uber." arXiv preprint arXiv:2004.03045 (2020), hereinafter referenced as PAN (NPL11 in Applicant’s 5/3/2023 IDS), in view of Last, Mark. "Online classification of nonstationary data streams” (2002), hereinafter referenced as LAST, and further in view of Reddi, Sashank, et al. "Doubly robust covariate shift correction." Proceedings of the AAAI conference on artificial intelligence. Vol. 29. No. 1. 2015, hereinafter referenced as REDDI (NPL4 in Applicant’s 10/17/2024 IDS), and further in view of Chen, Zhiqiang, et al. "A multi-level weighted concept drift detection method." The Journal of Supercomputing 79.5 (Sept. 2022), hereinafter referenced as CHEN.
Regarding Claim 1
PAN teaches:
A method of managing model drift in a machine learning model, the method comprising: (PAN, p. 3, section 3: “The feature importance and propensity score from the adversarial classifier can be used to detect concept drift between the training and test data, and provide insights on the cause of the concept drift such as which features and subsamples in the training data are most different from ones in the test data. In addition to concept drift detection, here, we propose three adversarial validation methods that address concept drift between the training and test data, and generate predictions adapted to the test dataset.”
PAN, p. 4, section 4: “Adversarial validation with three different methods, feature selection, validation selection, and inverse propensity weighting (IPW) are applied to seven datasets from AutoML3 for Lifelong Machine Learning Challenge as well as MaLTA dataset.”
Examiner’s Note: PAN discloses techniques for detecting and addressing concept drift with respect to machine learning models)
extracting test data and training data ...; (PAN, p. 3, section 3: “We start with a labeled training dataset ... and an unlabeled test dataset...”;
PAN, p. 3, section 3.2: “With validation data selection, we construct a new validation dataset ... by selecting from the training data so that the empirical distribution of the features data is similar to the test data.”;
PAN, p. 4, section 4: “GBDT models are trained with early stopping: 25% of the training
dataset is used as the validation set, and GBDT models train until the performance on the validation set stops improving”;
Examiner’s Note: PAN discloses selecting different validation, test, and training datasets from training data)
predicting (PAN, p. 2, section 2.2: “In causal inference, propensity score modeling addresses the heterogeneity between the treatment and control group data by training a classifier to predict if a sample belongs to a treatment group. Rosenbaum and Rubin argue in [12] that it is sufficient to achieve the balance in the distributions between the treatment and control groups by matching on the single dimensional propensity score alone, which is significantly more efficient than matching on the joint distribution of all confounding variables. ... With randomized controlled trial (RCT) data, we can estimate the average treatment effect (ATE) by calculating the difference between the average outcomes of the treatment and control groups, since the control group forms a valid counterfactual prediction given the similar distribution to the treatment group.”
PAN, p. 5, section 6: “adversarial feature selection, which is analogous to model selection and outlier removal in the regular machine learning, is more robust to the errors in estimating the distributional differences”;
Examiner’s Note: PAN discloses propensity score modeling that predicts treatments using both treatment and control group data, where the modeling is robust to errors)
measuring concept drift between the test data and the training data with respect to the machine learning model, ... (PAN, p. 3, section 3: “The feature importance and propensity score from the adversarial classifier can be used to detect concept drift between the training and test data, and provide insights on the cause of the concept drift such as which features and subsamples in the training data are most different from ones in the test data.”;
Examiner’s Note: PAN discloses detecting (corresponding to recited “measuring”) concept drift between the test data and training data
selecting ... feature vectors... (PAN, p. 3, section 3.1: “If distributions of the features from the train and test data are similar, we expect the adversarial classifier to be as good as random guesses. However, if the adversarial classifier can distinguish between training and test data well (i.e. AUC score ≫ 50%), the top features from the adversarial classifier are potential candidates exhibiting concept drift between the train and test data. We can then exclude these features from model training. Such feature selection can be automated by determining the number of features to exclude based on the performance of adversarial classifier (e.g. AUC score) and raw feature importance values (e.g. mean decrease impurity (MDI) in Decision Trees) as follows: (1) Train an adversarial classifier that predicts P({train, test }|X) to separate train and test.”;
Examiner’s Note: PAN discloses selecting feature vectors for the training and test datasets using an adversarial classifier)
retraining the machine learning model ... (PAN, p. 2, section 2.3: “models were retrained with new training data consisting of the old training and latest test data”)
However, PAN fails to explicitly teach:
extracting test data and training data from an ordered data stream, the test data being extracted from a detection window in the ordered data stream and the training data being extracted from a sliding reference window that precedes the detection window in the ordered data stream;
for the ordered data stream / of the ordered data stream
doubly robust
wherein the concept drift is measured as an expectation of differences between the doubly robust outcomes for the ordered data stream and the doubly robust outcomes for the training data;
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and
retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition.
However, in a related field of endeavor (a network that “adapts itself automatically to the rate of concept drift in a non-stationary data stream”, see p. 6, section 1), LAST teaches:
extracting test data and training data from an ordered data stream, the test data being extracted from a detection window in the ordered data stream and the training data being extracted from a sliding reference window that precedes the detection window in the ordered data stream; (LAST, p. 6, section 1: “Thus, OLIN can be applied to a time-changing data stream of arbitrary duration.”
LAST, p. 12, section 3.1: “The system repeatedly applies the IN algorithm to
a sliding window of training examples and it dynamically adapts the size of the training
window and the frequency of model re-construction to the current rate of concept drift.”;
LAST, p. 13, section 3.1: “In the sample data stream of Figure 2, each one of V0 examples in the validation interval [t2, t3] is classified by a model induced from T0 examples of the training interval [t0, t2]. The number of examples in the training and the validation intervals do not have to be equal. At the time t3 the network is re-constructed by the Learning Module using T1 examples from the training interval [t1, t3] and subsequently applied to V1 examples in the validation interval [t3, t4].”
LAST, p. 15, section 3.1: “Table 1 shows the pseudo-code outline of the OLIN algorithm. The algorithm does some initializations and then processes a user-specified number of incoming examples from a continuous data stream.”
PNG
media_image1.png
410
580
media_image1.png
Greyscale
Examiner’s Note: LAST teaches the OLIN algorithm that is applied to continuous, time-changing data streams (corresponding to recited “ordered data stream”, where the data stream is ordered with respect to time), where training data is selected, using a sliding window, from a training interview from t0 to t2 (corresponding to recited “sliding reference window”), which precedes the validation data window from time t2 to t3 (corresponding to recited “detection window in the ordered stream data”)
for the ordered data stream / of the ordered data stream (LAST, p. 6, section 1: “Thus, OLIN can be applied to a time-changing data stream of arbitrary duration.”
LAST, p. 15, section 3.1: “Table 1 shows the pseudo-code outline of the OLIN algorithm. The algorithm does some initializations and then processes a user-specified number of incoming examples from a continuous data stream.”
Examiner’s Note: LAST teaches the OLIN algorithm that is applied to continuous, time-changing data streams (corresponding to recited “ordered data stream”)
The combination of PAN and LAST makes obvious:
extracting test data and training data from an ordered data stream, the test data being extracted from a detection window in the ordered data stream and the training data being extracted from a sliding reference window that precedes the detection window in the ordered data stream (Examiner’s Note: the PAN-LAST combination now modifies the system of PAN to collect training and test/validation data from a continuous time-changing data stream, with different sliding and detection windows, as explained by LAST)
predicting for the ordered data stream based on a combination of treatments predicted based on controls of the ordered data stream and outcomes predicted based on the treatments and the controls of the ordered data stream, wherein the (Examiner’s Note: the PAN-LAST combination now modifies the system of PAN to collect training and test/validation data from a continuous time-changing data stream, with different sliding and detection windows, as explained by LAST, so that PAN can predict treatments based on treatment and control groups as disclosed by PAN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN and LAST as explained above. As explained by LAST, one of ordinary skill would have been motivated to do so because LAST recognizes that “for high-volume non-stationary data streams, where the actual rate of drift is unknown in advance, the run time of the algorithm may grow indefinitely.” (p. 3, section 1). Therefore, one of ordinary skill in the art would have been motivated to use the teachings of LAST to apply to a “time-changing data stream of arbitrary duration” because LAST teaches that the “cumulative accuracy of the models produced by OLIN tends to be higher than the accuracy obtained with a fixed-size sliding window” (p. 9, section 1).
However, PAN and LAST fail to explicitly teach:
doubly robust
wherein the concept drift is measured as an expectation of differences between the doubly robust outcomes for the ordered data stream and the doubly robust outcomes for the training data;
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and
retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition.
However, in a related field of endeavor (covariate shift with respect to real data, see p. 2949, Introduction section), REDDI teaches:
doubly robust (REDDI, p. 2950, Introduction section: “We develop a simple, yet powerful, framework for doubly robust estimation in the context of covariate shift correction”;
REDDI, p. 2953, Experiments section: “We present our empirical results in this section. We apply doubly robust covariate shift correction to a broad range of UCI datasets and a real-world dataset to demonstrate its performance.”
Examiner’s Note: REDDI discloses determining covariate shift (which para. 0001 in the instant specification explains is the same as “data drift”) along with doubly robust estimation)
wherein the concept drift is measured as an expectation of differences between the doubly robust outcomes for the ordered data stream and the doubly robust outcomes for the training data; (REDDI, p.2950, “Doubly Robust Covariate Shift Correction” section:
PNG
media_image2.png
380
484
media_image2.png
Greyscale
Examiner’s Note: REDDI discloses determining covariate shift (which para. 0001 in the instant specification explains is the same as “data drift”), with the goal of reducing the drift using a measurement based on the expectation of minimizing a loss function with respect to outcomes (y) based as a function of inputs (x) such as training data)
The combination of PAN, LAST, and REDDI makes obvious:
predicting doubly robust outcomes for the ordered data stream based on a combination of treatments predicted based on controls of the ordered data stream and outcomes predicted based on the treatments and the controls of the ordered data stream, wherein the doubly robust outcomes include doubly robust outcomes for the test data and doubly robust outcomes for the training data (Examiner’s Note: the PAN-LAST-REDDI combination now modifies PAN to utilize the doubly robust estimators of REDDI, as applied to the continuous data stream of LIFT, when predicting outcomes based on treatment and control groups as in PAN)
wherein the concept drift is measured as an expectation of differences between the doubly robust outcomes for the ordered data stream and the doubly robust outcomes for the training data (Examiner’s Note: the PAN-LAST-REDDI combination now modifies PAN to calculate concept drift in a manner similar to covariate shift as in REDDI, where such drift is measured to minimize a loss function (e.g., minimize the difference between outcomes for estimated predictions from the data stream vs. predictions based on training data), where such outputs are measured in a doubly robust manner as in REDDI)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, and REDDI as explained above. As explained by REDDI, one of ordinary skill would have been motivated to do so because REDDI’s “approach is particularly simple insofar as it requires essentially no additional code to use — all that is required in practice is to allow for reweighting and offset-correction in a linear model, a decision tree, or any other estimator that might be at hand. Of particular importance is the fact that we found our approach never to be worse than unweighted solution, something that cannot be said in general for covariate shift correction.” (p. 6, Conclusion section).
However, PAN, LAST, and REDDI fail to explicitly teach:
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and
retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition.
However, in a related field of endeavor (concept drift detection with respect to artificial intelligence models, see p. 5155, section 1), CHEN teaches:
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and (CHEN, p. 5156, section 2.1: “Suppose the data stream is in the form of consecutive (xt , yt) instances, where t = 1, 2, 3…., and xt is a feature vector obtained by the predictor based on the feature vector xt at a specific time can be denoted by ̂yt . Then, the concept drift in the time t0 to t1 can be defined as formula (1) [18]. Here, pt represents the joint probability distribution between the feature vector xt and the target class label yt at time t. The change of the data flow distribution is the concept drift, which can be reflected in the change of the joint probability distribution. and y belongs to a set with n class labels That is, y ∈ {y1, y2,⋯yn}.”
CHEN, p. 5163, section 3.1: “Finally, in the“drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded (corresponding to recited “retraining condition”), and further discloses obtaining feature vectors to perform training)
retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition. (CHEN, p. 5156, section 2.1: “Suppose the data stream is in the form of consecutive (xt , yt) instances, where t = 1, 2, 3…., and xt is a feature vector obtained by the predictor based on the feature vector xt at a specific time can be denoted by ̂yt . Then, the concept drift in the time t0 to t1 can be defined as formula (1) [18]. Here, pt represents the joint probability distribution between the feature vector xt and the target class label yt at time t. The change of the data flow distribution is the concept drift, which can be reflected in the change of the joint probability distribution. and y belongs to a set with n class labels That is, y ∈ {y1, y2,⋯yn}.”
CHEN, p. 5163, section 3.1: “Finally, in the“drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded (corresponding to recited “satisfy retraining condition”))
The combination of PAN, LAST, REDDI, and CHEN makes obvious:
selecting retraining feature vectors from feature vectors of the ordered data stream, based on the concept drift being measured to satisfy a retraining condition; and (Examiner’s Note: the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) from the continuous data stream of LAST (e.g., newer data) in order to retrain the model using newer data)
retraining the machine learning model using the retraining feature vectors, based on the concept drift being measured to satisfy the retraining condition. (Examiner’s Note: the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) from the continuous data stream of LAST (e.g., newer data) in order to actually retrain the model using newer data as in both PAN and CHEN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, and CHEN as explained above. As explained by CHEN, one of ordinary skill would have been motivated to do so because CHEN “solve[s] the problem of the inability of existing drift detection methods to balance the detection delay, false positives, false negatives, and space–time efficiency when detecting abrupt and gradual concept drift.” (p. 5155, section 1). One of ordinary skill would further understand the benefit of limiting the retraining to only when actual shift is detected because retraining a machine learning model is time-intensive and requires a lot of resource consumption.
Regarding Claim 4
PAN, LAST, REDDI, and CHEN teach the method of claim 1 as explained above. PAN further teaches and makes obvious:
selecting (PAN, p. 3, section 3.1: “If distributions of the features from the train and test data are similar, we expect the adversarial classifier to be as good as random guesses. However, if the adversarial classifier can distinguish between training and test data well (i.e. AUC score ≫ 50%), the top features from the adversarial classifier are potential candidates exhibiting concept drift between the train and test data. We can then exclude these features from model training. Such feature selection can be automated by determining the number of features to exclude based on the performance of adversarial classifier (e.g. AUC score) and raw feature importance values (e.g. mean decrease impurity (MDI) in Decision Trees) as follows: (1) Train an adversarial classifier that predicts P({train, test }|X) to separate train and test.”;
Examiner’s Note: PAN discloses selecting feature vectors for the training and test datasets using an adversarial classifier)
However, PAN, LAST, and REDDI fail to explicitly teach:
the retraining feature vectors
However, in a related field of endeavor (concept drift detection with respect to artificial intelligence models, see p. 5155, section 1), CHEN teaches and makes obvious:
selecting the retraining feature vectors based on scores generated by an adversarial feature classifier; and (CHEN, p. 5163, section 3.1: “Finally, in the “drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded; the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) using the adversarial classifier of PAN order to actually retrain the model using newer data as in both PAN and CHEN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, and CHEN as explained above. As explained by CHEN, one of ordinary skill would have been motivated to do so because CHEN “solve[s] the problem of the inability of existing drift detection methods to balance the detection delay, false positives, false negatives, and space–time efficiency when detecting abrupt and gradual concept drift.” (p. 5155, section 1). One of ordinary skill would further understand the benefit of limiting the retraining to only when actual shift is detected because retraining a machine learning model is time-intensive and requires a lot of resource consumption.
Regarding Claim 5
PAN, LAST, REDDI, and CHEN teach the method of claim 4 as explained above. PAN further teaches and makes obvious:
selecting (PAN, p. 3, section 3.1: “If distributions of the features from the train and test data are similar, we expect the adversarial classifier to be as good as random guesses. However, if the adversarial classifier can distinguish between training and test data well (i.e. AUC score ≫ 50%), the top features from the adversarial classifier are potential candidates exhibiting concept drift between the train and test data. We can then exclude these features from model training. Such feature selection can be automated by determining the number of features to exclude based on the performance of adversarial classifier (e.g. AUC score) and raw feature importance values (e.g. mean decrease impurity (MDI) in Decision Trees) as follows: (1) Train an adversarial classifier that predicts P({train, test }|X) to separate train and test. (2) If the AUC score of the adversarial classifier is greater than an AUC threshold θauc , remove features ranked within top x% of remaining features in feature importance ranking and with raw feature importance values higher than a threshold θimp (e.g. MDI > 0.1). (3) Go back to Step 1, if AUC score greater than θauc . (4) Once the adversarial AUC drops lower than θauc , train an outcome classifier with the selected features and original target variable.”;
Examiner’s Note: PAN discloses selecting feature vectors for the training and test datasets using an adversarial classifier, where an AUC score and an AUC threshold (θauc) (corresponding to recited “AUC hyperparameter”) are used when selecting feature vectors)
However, PAN, LAST, and REDDI fail to explicitly teach:
the retraining feature vectors
However, in a related field of endeavor (concept drift detection with respect to artificial intelligence models, see p. 5155, section 1), CHEN teaches and makes obvious:
selecting the retraining feature vectors based on an area-under-the-curve (AUC) score generated by the adversarial feature classifier and an AUC condition hyperparameter (CHEN, p. 5163, section 3.1: “Finally, in the “drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded; the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) using the adversarial classifier of PAN order to actually retrain the model using newer data as in both PAN and CHEN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, and CHEN as explained above. As explained by CHEN, one of ordinary skill would have been motivated to do so because CHEN “solve[s] the problem of the inability of existing drift detection methods to balance the detection delay, false positives, false negatives, and space–time efficiency when detecting abrupt and gradual concept drift.” (p. 5155, section 1). One of ordinary skill would further understand the benefit of limiting the retraining to only when actual shift is detected because retraining a machine learning model is time-intensive and requires a lot of resource consumption.
Regarding Claim 6
PAN, LAST, REDDI, and CHEN teach the method of claim 4 as explained above. PAN further teaches and makes obvious:
selecting (PAN, p. 3, section 3.1: “If distributions of the features from the train and test data are similar, we expect the adversarial classifier to be as good as random guesses. However, if the adversarial classifier can distinguish between training and test data well (i.e. AUC score ≫ 50%), the top features from the adversarial classifier are potential candidates exhibiting concept drift between the train and test data. We can then exclude these features from model training. Such feature selection can be automated by determining the number of features to exclude based on the performance of adversarial classifier (e.g. AUC score) and raw feature importance values (e.g. mean decrease impurity (MDI) in Decision Trees) as follows: (1) Train an adversarial classifier that predicts P({train, test }|X) to separate train and test. (2) If the AUC score of the adversarial classifier is greater than an AUC threshold θauc , remove features ranked within top x% of remaining features in feature importance ranking and with raw feature importance values higher than a threshold θimp (e.g. MDI > 0.1). (3) Go back to Step 1, if AUC score greater than θauc . (4) Once the adversarial AUC drops lower than θauc , train an outcome classifier with the selected features and original target variable.”;
Examiner’s Note: PAN discloses selecting feature vectors for the training and test datasets using an adversarial classifier, where raw importance feature scores and an importance value threshold (θimp) (corresponding to recited “raw feature importance condition hyperparameter”) are used when selecting feature vectors)
However, PAN, LAST, and REDDI fail to explicitly teach:
the retraining feature vectors
However, in a related field of endeavor (concept drift detection with respect to artificial intelligence models, see p. 5155, section 1), CHEN teaches and makes obvious:
selecting the retraining feature vectors based on a raw feature importance value and a raw feature importance condition hyperparameter. (CHEN, p. 5163, section 3.1: “Finally, in the “drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded; the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) using the adversarial classifier of PAN order to actually retrain the model using newer data as in both PAN and CHEN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, and CHEN as explained above. As explained by CHEN, one of ordinary skill would have been motivated to do so because CHEN “solve[s] the problem of the inability of existing drift detection methods to balance the detection delay, false positives, false negatives, and space–time efficiency when detecting abrupt and gradual concept drift.” (p. 5155, section 1). One of ordinary skill would further understand the benefit of limiting the retraining to only when actual shift is detected because retraining a machine learning model is time-intensive and requires a lot of resource consumption.
Regarding Claim 8
LAST teaches:
A computing system (LAST, p. 16: “This imposes an additional limitation on the size of the training window, which can be handled by a given computer system.”;
LAST, p. 22: “All runs were carried out on a Pentium III processor with 128 MB of RAM.”)
PAN teaches:
for managing model drift in a machine learning model, the computing system comprising: (PAN, p. 3, section 3: “The feature importance and propensity score from the adversarial classifier can be used to detect concept drift between the training and test data, and provide insights on the cause of the concept drift such as which features and subsamples in the training data are most different from ones in the test data. In addition to concept drift detection, here, we propose three adversarial validation methods that address concept drift between the training and test data, and generate predictions adapted to the test dataset.”
PAN, p. 4, section 4: “Adversarial validation with three different methods, feature selection, validation selection, and inverse propensity weighting (IPW) are applied to seven datasets from AutoML3 for Lifelong Machine Learning Challenge as well as MaLTA dataset.”
Examiner’s Note: PAN discloses techniques for detecting and addressing concept drift with respect to machine learning models)
The remaining limitations in claim 8 correspond to the method of claim 1 and are therefore rejected for the same reasons explained above with respect to claim 1.
Claim 11 depends from claim 8 and claims a computing system that corresponds to the method of claim 2, and is therefore rejected for the same reasons explained above with respect to claims 4 and 8.
Claim 12 depends from claim 11 and claims a computing system that corresponds to the method of claim 5, and is therefore rejected for the same reasons explained above with respect to claims 5 and 11.
Claim 13 depends from claim 11 and claims a computing system that corresponds to the method of claim 5, and is therefore rejected for the same reasons explained above with respect to claims 5 and 11.
Claim 18 depends from claim 15 and claims a tangible processor-readable storage media that corresponds to the method of claim 4, and is therefore rejected for the same reasons explained above with respect to claims 4 and 15.
Claim 19 depends from claim 18 and claims a tangible processor-readable storage media that corresponds to the method of claim 5, and is therefore rejected for the same reasons explained above with respect to claims 5 and 18.
Claims 2-3, 9-10, and 16-17 are rejected under 35 U.S.C. 103 as being unpatentable over PAN, in view of LAST, REDDI, and CHEN, and further in view of Allan, Victoria, et al. "Propensity score matching and inverse probability of treatment weighting to address confounding by indication in comparative effectiveness research of oral anticoagulants." Journal of comparative effectiveness research 9.9 (2020): 603-614, hereinafter referenced as ALLAN.
Regarding Claim 2
PAN, LAST, REDDI, and CHEN teach the method of claim 1 as explained above. PAN further teaches and makes obvious:
wherein the training data includes observed outcomes, and the (PAN, p. 2, section 2.2: “In causal inference, propensity score modeling addresses the heterogeneity between the treatment and control group data by training a classifier to predict if a sample belongs to a treatment group. Rosenbaum and Rubin argue in [12] that it is sufficient to achieve the balance in the distributions between the treatment and control groups by matching on the single dimensional propensity score alone, which is significantly more efficient than matching on the joint distribution of all confounding variables. ... With randomized controlled trial (RCT) data, we can estimate the average treatment effect (ATE) by calculating the difference between the average outcomes of the treatment and control groups, since the control group forms a valid counterfactual prediction given the similar distribution to the treatment group.”
PAN, p. 3, section 3.2: “With validation data selection, we construct a new validation dataset ... by selecting from the training data so that the empirical distribution of the features data is similar to the test data.”;
PAN, p. 5, section 6: “adversarial feature selection, which is analogous to model selection and outlier removal in the regular machine learning, is more robust to the errors in estimating the distributional differences”;
Examiner’s Note: PAN discloses propensity score modeling that predicts treatments using both treatment and control group data, where the modeling is robust to errors, and that training data is collected to train the models)
However, PAN and LAST fail to explicitly teach:
doubly robust
added to an inverse propensity weighting of residuals between the observed outcomes and the outcomes predicted based on the treatments and the controls of the ordered data stream.
However, in a related field of endeavor (covariate shift with respect to real data, see p. 2949, Introduction section), REDDI teaches and makes obvious:
doubly robust (REDDI, p. 2950, Introduction section: “We develop a simple, yet powerful, framework for doubly robust estimation in the context of covariate shift correction”;
REDDI, p. 2953, Experiments section: “We present our empirical results in this section. We apply doubly robust covariate shift correction to a broad range of UCI datasets and a real-world dataset to demonstrate its performance.”
Examiner’s Note: REDDI discloses determining covariate shift (which para. 0001 in the instant specification explains is the same as “data drift”) along with doubly robust estimation; the PAN-LAST-REDDI combination now modifies PAN where outputs are measured in a doubly robust manner as in REDDI)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, and REDDI as explained above. As explained by REDDI, one of ordinary skill would have been motivated to do so because REDDI’s “approach is particularly simple insofar as it requires essentially no additional code to use — all that is required in practice is to allow for reweighting and offset-correction in a linear model, a decision tree, or any other estimator that might be at hand. Of particular importance is the fact that we found our approach never to be worse than unweighted solution, something that cannot be said in general for covariate shift correction.” (p. 6, Conclusion section).
However, PAN, LAST, REDDI, and CHEN fail to explicitly teach:
added to an inverse propensity weighting of residuals between the observed outcomes and the outcomes predicted based on the treatments and the controls of the ordered data stream.
However, in a related field of endeavor (treatment and control groups, see p. 604), ALLAN teaches and makes obvious:
added to an inverse propensity weighting of residuals between the observed outcomes and the outcomes predicted based on the treatments and the controls of the ordered data stream. (ALLAN, p. 604: “Propensity score matching(PSM) and inverse probability of treatment weighting (IPTW) are increasingly popular methods used to address confounding by indication in RWE studies. ... While PSM and IPTW endeavor to achieve the same objective in balancing out differences between treatment groups, the two methods provide a different measurement of the treatment effect and this should be interpreted accordingly. When applied to the same data, PSM and IPTW may not always point to the same findings suggesting that these methods are not strictly interchangeable.”
ALLAN, p. 605: “Four case studies involving the use of PSM, IPTW or a combination of both methods in real-world CER of OACs will be presented throughout this article to demonstrate how these methods have been applied in practical terms.”;
Examiner’s Note: the PAN-LAST-REDDI-CHEN-ALLAN combination now modifies the feature vector selection of PAN to utilize the IPTW weighting method, in combination with the PSM method already taught by PAN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, CHEN, and ALLAN as explained above. As explained by ALLAN, one of ordinary skill would have been motivated to do so because ALLAN teaches that PSM and IPTW are not “strictly interchangeable” and that a combination of the methods may deliver better results. (pp. 604-605). One of ordinary skill would also understand the benefit of using a combination of both methods in view of other peer-reviewed studies using a similar combination.
Regarding Claim 3
PAN, LAST, REDDI, CHEN, and ALLAN teach the method of claim 2 as explained above. However, PAN, LAST, REDDI, and CHEN fail to explicitly teach:
wherein the inverse propensity weighting of the residuals is based on the treatments predicted based on the controls of the ordered data stream.
However, in a related field of endeavor (treatment and control groups, see p. 604), ALLAN teaches and makes obvious:
wherein the inverse propensity weighting of the residuals is based on the treatments predicted based on the controls of the ordered data stream. (ALLAN, p. 610: “In the IPTW method, weights are assigned to patients based on the inverse of their probability of receiving treatment, as estimated by the propensity score. IPTW results in a pseudo-population in which patients with a high probability of receiving treatment have a smaller weight and patients with a low probability of receiving treatment have a larger weight and thus the distribution of measured patient characteristics used to calculate the propensity score becomes independent of treatment assignment. IPTW provides an estimation of the ATE, because the study population is re-weighted to assess the effects of the treatment in the scenario that it was offered to all patients within the population”
Examiner’s Note: the PAN-LAST-REDDI-CHEN-ALLAN combination now modifies the feature vector selection of PAN to utilize the IPTW weighting method as applied to treatments, where such treatments are utilized with respect to a control group as in PAN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, CHEN, and ALLAN as explained above. As explained by ALLAN, one of ordinary skill would have been motivated to do so because ALLAN teaches that PSM and IPTW are not “strictly interchangeable” and that a combination of the methods may deliver better results. (pp. 604-605). One of ordinary skill would also understand the benefit of using a combination of both methods in view of other peer-reviewed studies using a similar combination.
Claim 9 depends from claim 8 and claims a computing system that corresponds to the method of claim 2, and is therefore rejected for the same reasons explained above with respect to claims 2 and 8.
Claim 10 depends from claim 9 and claims a computing system that corresponds to the method of claim 3, and is therefore rejected for the same reasons explained above with respect to claims 3 and 9.
Claim 16 depends from claim 15 and claims a tangible processor-readable storage media that corresponds to the method of claim 2, and is therefore rejected for the same reasons explained above with respect to claims 2 and 15.
Claim 17 depends from claim 16 and claims a tangible processor-readable storage media that corresponds to the method of claim 3, and is therefore rejected for the same reasons explained above with respect to claims 3 and 16.
Claims 7, 14, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over PAN, in view of LAST, REDDI, and CHEN, and further in view of Altmann, André, et al. "Permutation importance: a corrected feature importance measure." Bioinformatics 26.10 (2010): 1340-1347, hereinafter referenced as ALTMANN.
Regarding Claim 7
PAN, LAST, REDDI, and CHEN teach the method of claim 4 as explained above. PAN further teaches and makes obvious:
selecting (PAN, p. 3, section 3.1: “If distributions of the features from the train and test data are similar, we expect the adversarial classifier to be as good as random guesses. However, if the adversarial classifier can distinguish between training and test data well (i.e. AUC score ≫ 50%), the top features from the adversarial classifier are potential candidates exhibiting concept drift between the train and test data. We can then exclude these features from model training. Such feature selection can be automated by determining the number of features to exclude based on the performance of adversarial classifier (e.g. AUC score) and raw feature importance values (e.g. mean decrease impurity (MDI) in Decision Trees) as follows: (1) Train an adversarial classifier that predicts P({train, test }|X) to separate train and test. (2) If the AUC score of the adversarial classifier is greater than an AUC threshold θauc , remove features ranked within top x% of remaining features in feature importance ranking and with raw feature importance values higher than a threshold θimp (e.g. MDI > 0.1). (3) Go back to Step 1, if AUC score greater than θauc . (4) Once the adversarial AUC drops lower than θauc , train an outcome classifier with the selected features and original target variable.”;
Examiner’s Note: PAN discloses selecting feature vectors for the training and test datasets using an adversarial classifier, where raw importance feature scores and an importance value threshold (θimp) (corresponding to recited “raw feature importance condition hyperparameter”) are used when selecting feature vectors)
However, PAN, LAST, and REDDI fail to explicitly teach:
the retraining feature vectors ...
raw permutation feature importance value and a raw permutation feature importance condition hyperparameter
However, in a related field of endeavor (concept drift detection with respect to artificial intelligence models, see p. 5155, section 1), CHEN teaches and makes obvious:
selecting the retraining feature vectors based on a raw (CHEN, p. 5163, section 3.1: “Finally, in the “drift level,” MWDDM_H and MWDDM_M are determined by the Hoeffding and Mcdiarmid bounds generated by the Hoeffding inequality and Mcdiarmid inequality, respectively. If the difference between the maximum weighted average of correct prediction and the weighted average of correct prediction in the long and short sliding windows is greater than the pre-defined threshold, the occurrence of a concept drift will be reported. At this time, the classifier will be reset to retrain to adapt to the new data distribution.”;
CHEN, p. 5165, section 3.2: “Lines 24–29 of the algorithm indicate that, during the "drift level,"MWDDM_H and MWDDM_M will determine whether Δs and Δl are greater than
a pre-defined threshold, and if so, will report the occurrence of a drift and reset the classifier for retraining.”;
Examiner’s Note: CHEN discloses retraining AI models for concept drift when a pre-defined threshold is exceeded; the PAN-LAST-REDDI-CHEN combination now modifies PAN to utilize the pre-defined threshold to determine if retraining is required as in CHEN, and based on such pre-defined threshold being exceeded, obtain new feature vectors (as in CHEN) using the adversarial classifier of PAN order to actually retrain the model using newer data as in both PAN and CHEN)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, and CHEN as explained above. As explained by CHEN, one of ordinary skill would have been motivated to do so because CHEN “solve[s] the problem of the inability of existing drift detection methods to balance the detection delay, false positives, false negatives, and space–time efficiency when detecting abrupt and gradual concept drift.” (p. 5155, section 1). One of ordinary skill would further understand the benefit of limiting the retraining to only when actual shift is detected because retraining a machine learning model is time-intensive and requires a lot of resource consumption.
However, PAN, LAST, REDDI, and CHEN fail to explicitly teach:
raw permutation feature importance value and a raw permutation feature importance condition hyperparameter
However, in a related field of endeavor (evaluating features for importance, see p. 1340, section 1), ALTMANN teaches:
raw permutation feature importance value (ALTMANN, p. 1341, section 1: “In this article, we introduce a heuristic for correcting biased measures of feature importance, called permutation importance (PIMP). The method normalizes the biased measure based on a permutation test and returns significance P-values for each feature. To preserve the relations between features, we use permutations of the outcome. We show that this method can be used to correct for the bias of feature importance computed with RF and MI. Moreover, our method can be used together with any learning method that assesses feature relevance, providing significance P-values for each predictor variable.”)
The PAN, LAST, REDDI, CHEN, and ALTMANN combination makes obvious:
selecting the retraining feature vectors based on a raw permutation feature importance value and a raw permutation feature importance condition hyperparameter.
(Examiner’s Note: As explained with respect to claim 6, the PAN-LAST-REDDI-CHEN combination teaches selecting retraining feature vectors based on a raw feature importance value and a raw permutation feature importance condition hyperparameter; the PAN-LAST-REDDI-CHEN-ALTMANN combination now modifies the “a raw feature importance value” of PAN to utilize the PIMP value of ALTMANN and for the importance hyperparameter threshold of PAN to be modified to apply to the PIMP value of ALTMANN).
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of PAN, LAST, REDDI, CHEN, and ALTMANN as explained above. As explained by ALTMANN, one of ordinary skill would have been motivated to do so because ALTMANN teaches that the PIMP metrics can be used to “correct[] biased measures of feature importance.” (p. 1341, section 1).
Claim 14 depends from claim 11 and claims a computing system that corresponds to the method of claim 7, and is therefore rejected for the same reasons explained above with respect to claims 7 and 11.
Claim 20 depends from claim 18 and claims a computing system that corresponds to the method of claim 7, and is therefore rejected for the same reasons explained above with respect to claims 7 and 18.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Zhu, Qun, et al. "A double-window-based classification algorithm for concept drifting data streams." 2010 IEEE International Conference on Granular Computing. IEEE, 2010. Discloses a double window mechanism for data streams. (p. 1, section 1).
Nigenda, David, et al. "Amazon sagemaker model monitor: A system for real-time insights into deployed machine learning models." Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2022. Describes a system architecture provided by Amazon to enable model owners to take corrective action to their models in view of “data, concept, bias, and feature attribution drift.” (Abstract).
US 11983742 B1 (Ma). “As applied to the systems and methods described herein, an indication that a user identifier is not associated with particular data may be used as a treatment assignment. The non-treated outcome in the treatment group may be counterfactual attributed conversions based on id-less data. DML models may incorporate advanced machine learning algorithms such as deep neural nets to improve accuracy. DML may involve two first-stage nuisance variable predictions: propensity score prediction and outcome predictions. The former may be used to re-weight outcome prediction using observations from a control group and the latter is used to improve outcome prediction.” (col. 4, lines 6-17).
US 20200380417 A1 (Briancon). “In some embodiments, OQM attributes may include count, unique count, null count, mean, max, min, standard-deviation, median, missing data source, new data source, missing data element, new data element, sparsity of reward, data type change, Accuracy, Precision, Recall, F1, ROC AUC, TPR, TNR, 1-FPR, 1-FNR, brier gain, 1-KS, lift statistic, CV Area under Curve, 1-CV turn on, CV plateau, 1-brier turn on, brier plateau, MAPk, TkCA, Coverage, entropy coverage, MAPk cohort, TkCA population, action percentage change, no action percentage change, action frequency rate, action recency rate, normalized action frequency rate, normalized action recency rate, expected reward, direct method, inverse propensity method, doubly robust method, weighted doubly robust, sequential doubly robust, magic doubly robust, incremental response rate, net incremental revenue, Mann Whitney Wilcoxon's U test, decile analysis test, effect size test, and economic efficiency.” (para. 0108).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MICHAEL C LEE whose telephone number is (571)272-4933. The examiner can normally be reached M-F 12:00 pm - 8:00 pm 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, Omar Fernandez Rivas can be reached at 571-272-2589. 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.
/MICHAEL C. LEE/Examiner, Art Unit 2128
Prosecution Insights