Detailed Action
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Acknowledgments
The RCE filed on 09/02/2026 is acknowledged.
Status of Claims
Claims 22-41 are pending.
In the Amendment filed on 08/14/2026 (entered via the RCE), claims 22, 32 and 41 were amended, and no claims were cancelled or added.
Claims 22-41 are rejected.
Response to Arguments
Regarding the rejection under 35 U.S.C. 101
Applicant’s arguments have been fully considered but are not persuasive.
The Office responds to Applicant’s arguments below. In the discussion below, headings and page numbers refer to Applicant’s Response, unless otherwise indicated.
A. Step 1 (p. 10)
The Examiner agrees with Applicant that the claims fall within the statutory categories under 35 U.S.C. 101.
B. Step 2A (p. 10)
1. Prong One (pp. 11-13)
Applicant cites the generating, retaining, training and tuning steps of claim 22, arguing that they are not directed to organizing human activity, that they do not merely use machine learning (ML) as a tool but recite specific operations of ML, that they cannot be performed in the human mind, and that they improve ML.
The Office respectfully disagrees.
First, some of the argued limitations are part of the abstract idea (namely, the bolded portions in the rejection in the body of the Office Action hereinbelow). Second, the limitations that are not part of the abstract idea (namely, the non-bolded portions in the rejection in the body of the Office Action hereinbelow), i.e., the ML model executed by the processor, the training, and the tuning step,1 are recited at a high level of generality, and not described: their relationship to the abstract idea is that they simply apply the abstract idea and/or generally link the use of the abstract idea to a particular technological environment or field of use -- in either case merely using ML in its ordinary capacity, not improving it -- and thus do not integrate the abstract idea into a practical application. Specifically, training an ML model and optimizing hyperparameters to maximize performance are standard ML operations (see MPEP 2106.05(f) "(2) … invokes computers or other machinery merely as a tool to perform an existing process") and do not amount to an improvement in ML. The presence of additional elements that amount to 'apply it' or generally link the use of the abstract idea to a particular technological environment or field of use does not detract from the presence of the abstract idea. Third, the claims were not rejected on the ground that they cannot be performed in the mind.
Further, the Office notes that, as per the rejection, steps such as receiving a processed action and a supplemental input associated with a transaction channel and including a chronological listing of all transactions relating to a current account of the user; generating a risk indicator based on the processed action and the supplemental input; responsive to a determination that the risk indicator exceeds a predetermined threshold, generating an alert indicating a probability of an unauthorized action; and selectively stopping, flagging for analyst review, or allowing execution of the processed action based on the risk indicator recite an abstract idea, e.g., determining the likelihood of fraud and preventing fraud.
2. Prong Two (pp. 13-14)
Applicant again argues that the above-indicated subject matter of claim 22 improves ML. This argument has been addressed above. To elaborate on the specific point raised here by Applicant, the use of feedback, including information/results from prior operation of the model, is a basic aspect of ML, not an improvement in ML.
Applicant also argues analogy to Example 39 of the USPTO Subject Matter Eligibility Guidance. The Office notes that the claim of Example 39 is directed to a “computer-implemented method of training a neural network for facial detection” -- subject matter entirely different from Applicant's instant claims, and further that the claim of Example 39 was found not to include a judicial exception. In contrast, Applicant’s independent claims do recite a judicial exception, as indicated above. Further, the ML training recited in Applicant's claims is not similar or analogous to that of the claim of Example 39.
Applicant also argues eligibility based on the 2024 AI Subject Matter Eligibility Examples (Examples 47-48). The Office notes that Examples 47-48 are directed to subject matter entirely different from Applicant's instant claims, and that the ML training recited in Examples 47-48 is not similar or analogous to that of Applicant's claims. Applicant's high-level, generic assertion of comparability is not persuasive.
C. Step 2B (pp. 14-15)
Applicant argues the same claimed subject matter as previously presented. This subject matter has already been addressed above and found to amount to an abstract idea and additional elements that merely apply the abstract idea and/or generally link the use of the abstract idea to a particular technological environment or field of use and as such cannot provide significantly more under Step 2B. MPEP 2106.05.I. Again, the ML limitations are merely standard ML operations (see MPEP 2106.05(f) "(2) … invokes computers or other machinery merely as a tool to perform an existing process") and do not amount to an improvement in ML.
Regarding the rejections under 35 U.S.C. 103
Applicant’s arguments have been fully considered but are either not persuasive or are moot based on newly cited portions of the existing prior art references used in the instant rejections.
The Office responds to Applicant's arguments below.
At page 16, Applicant discusses Comeaux. Although not quoting the previous Office Action precisely, Applicant notes that in the previous Office Action Comeaux was not cited as teaching the entirety of the previous version of the retaining step.
At pp. 16-17, Applicant cites the last three steps of claim 22 (retaining step, training step, and tuning step), which encompass the instant amendments and pertain to training and tuning a machine learning model.
At pp. 17-18, Applicant argues that Shevyrev does not teach the last three steps of claim 22.
Initially, it is noted that Comeaux is now cited as teaching the entirety of the instantly amended retaining step (which has been shortened compared to the previous version), based on and consistent with teachings cited in the previous Office Action (citations and explanation are provided in the rejection hereinbelow). Comeaux is also now cited as teaching the newly added training step, based on newly cited portions of Comeaux (citations and explanation are provided in the rejection hereinbelow). Applicant does not present substantive argument to the effect that Comeaux does not teach these steps.
Secondly, Shevyrev is now cited as teaching all of the last 3 steps of claim 22, based on and consistent with teachings cited in the previous Office Action but also based on newly cited portions of Shevyrev in order to further flesh out how the pre-existing claimed subject matter is taught and in view of the claimed subject matter newly added in the instant amendment (citations and explanation are provided in the rejection hereinbelow). Full explanation of how Shevyrev teaches the last three steps of claim 22 is provided in the rejection hereinbelow, and this explanation addresses Applicant's arguments against Shevyrev.
In addition, the Office notes that Applicant's arguments against Shevyrev rely in part on subject matter not claimed. For clarity, Applicant's argument, in pertinent part, is set forth below:
However, Shevyrev does not disclose the claimed sequence of "retaining...information associated with one or more previously generated risk indicators," subsequently "receiving additional information...including additional transactional data, customer characteristics, and historical data," and performing "tuning based on the training." Rather, the portions relied upon by the Office describe model-training operations involving predictions, evaluation of outcomes, and parameter optimization. Office Action at 17-18. Shevyrev does not disclose "tuning" that occurs based on "training the machine learning model based on additional information" after retention of information associated with previously generated risk indicators, as required by amended claim 22. (p. 17; emphasis added)
As per Applicant's argument above, Applicant argues in part based on an alleged claimed sequence / chronological order of operations -- as reflected by the underlined words above. However, this temporal language ("sequence," "subsequently," "after") is not recited in the claims. While the tuning and training steps recite information recited in previously listed steps, this fact does not require the sequence / chronological order that Applicant argues, at least not the entirety of it. For example, the claim does not require that the training step occur after the retaining step. In this regard, note that the "additional information" of the training step may be identified without the retaining step having been performed. Specifically, the "additional information" is recited as "including at least one of [1] additional transactional data, [2] customer characteristics, and [3] historical data associated with the previously generated risk indicators" – these three types/items of information (namely, [1], [2], and [3]) can be identified without the retaining step having been performed, e.g., without the risk indicator (i.e., the output of the machine learning model2) having been generated/identified.3
In addition to the unclaimed sequence / chronological order of operations, Applicant also argues a step of "receiving additional information...," which is not claimed.
At pp. 19-20, Applicant argues against the outstanding rejection of claims 25, 26, 35 and 36, which is based on prior art and legal precedent. Applicant's argument is moot in view of the revised rejection of those claims, which relies on newly cited portions of Comeaux and does not rely on legal precedent. Citations and explanation are provided in the rejection hereinbelow.
Claim Interpretation
Claims 22, 32 and 41 recite:
receiving, by the at least one processor, from at least one external database, at least one supplemental input associated with at least one of the transaction channel, the processed action, or the user, wherein the at least one supplemental input includes historical data including a chronological listing of all transactions relating to a current account of the user;
The term "supplemental" is a relative term (i.e., supplemental to what?). In light of specification ¶ 0043 and the preceding step of claims 22, 32 and 41, it is understood that the recited "supplemental input" is intended to be supplemental to the processed action received in the preceding claim step. Notably, there is only one step preceding this step of receiving the supplemental input, and that preceding step recites receipt of a processed action, which indeed serves as an input to the machine learning model. (As per the generating step, and the corresponding portion of the specification and figures,4 the processed action and the supplemental input are the inputs to the machine learning model.) Therefore, it may seem to follow naturally that the previously recited processed action is the only item that the supplemental input could be supplemental to. Accordingly, it is believed that one of ordinary skill in the art would understand what the supplemental input is relative to, and as such the term "supplemental" is not found to be indefinite. MPEP 2173.05(b) ("The use of relative terminology in claim language … does not automatically render the claim indefinite under 35 U.S.C. 112(b) …. Acceptability of the claim language depends on whether one of ordinary skill in the art would understand what is claimed, in light of the specification."). For purposes of examination, the recited "supplemental input" is interpreted as explained here.
Likewise, claims 23, 31 and 33 also recite "the at least one supplemental input" of base claims 22 and 32. Inasmuch as this recitation refers to the same recitation in the base claim, it is interpreted in the same way as the recitation of the base claim is interpreted, as per above.
Claims 22, 32 and 41 recite:
tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model.
On its face the above-quoted tuning step might be understood as saying that the tuning of the risk indicator serves to identify the optimal hyperparameter values. Such an interpretation would not be consistent with the knowledge/understanding of one of ordinary skill in the art.
Specifically, in the art of machine learning, a model may be tuned by tuning or optimizing the values of hyperparameters in order to improve model performance. (Model performance may be improved in various respects, for example, in the respect of increasing the accuracy of the model's results/outputs (e.g., by minimizing the loss function).) The optimizing of hyperparameter values is conventionally performed by one of three methods: grid search, random search, or Bayesian optimization. Grid search is a brute force method that tries all possible combinations of parameters within a defined range. Random search randomly selects only a portion of the parameters from a defined range. Bayesian optimization optimizes its parameter selection in each round, using a probabilistic model based on the results of previous rounds (instead of random selection).5 None of these methods involves tuning the output of the model. Accordingly, it is not clear how optimizing the hyperparameter values could be achieved by tuning the output of the model or, in other words, it is not clear how "tuning the risk indicator"6 could "identify optimal values of one or more hyperparameters to maximize performance of the machine learning model."
Moreover, we note that, in view of Applicant's disclosure, Applicant's teachings in regard to the claimed tuning step are understood to be in accordance with the knowledge/understanding of one of ordinary skill in the art. See Applicant's specification ¶ 0071:
In some embodiments, the machine learning model is trained to retain information associated with one or more previously generated indicators to tune a currently generated indicator. The information may be used to tune the information used to generate the previous indicator. [1] Tuning as used herein may refer to an experimental process of finding optimal values of hyperparameters to maximize model performance. [2] As used herein and generally understood in the art, hyperparameters are values selected to control a learning process of a model. The tuning occurs once the model receives relevant additional information that may include transactional data, customer characteristics, and historical data to optimize the performance of the model to correctly detect fraudulent transactions. (emphasis added; bracketed numerals added for reference in the discussion below)
Regarding [1], this statement is consistent with the knowledge/understanding of one of ordinary skill in the art. Regarding [2], this statement is deemed to reflect Applicant's intent that its teachings in this regard are in conformity with the knowledge/ understanding of one of ordinary skill in the art.
Therefore, we believe the disclosure reflects an intent on the part of Applicant to the effect that its teachings in this regard are in accordance with the knowledge/ understanding of one of ordinary skill in the art.
In addition, we are guided by the principle of charity: where language appears to be ambiguous or unclear, we generously interpret it in the most sensible/reasonable and constructive way, assuming the speaker (here, the drafter) intended to make sense.7 Rather than sticking to an interpretation that, although consistent with the literal wording of the claim, is non-sensical or implausible in light of knowledge of the art, we believe that, all other things being equal, adopting the principle of charity promotes compact prosecution.
Moreover, all other things are not equal: as mentioned above, Applicant demonstrates an intent to be in conformity with the general knowledge/understanding of the art. This factor further suggests that adopting the principle of charity and interpreting the claim language as being consistent with the general knowledge/understanding of the art promotes compact prosecution.
To be sure, the claim language "tuning the risk indicator" may arguably suggest that Applicant is attempting to be its own lexicographer and in this respect possibly attempting to convey something beyond the general knowledge/understanding of the art. Specifically, it is not ordinary parlance in the art to speak of "tuning" an output8 of a machine learning model. Rather, as explained above, the art generally speaks of tuning a machine learning model, or tuning the hyperparameters (or, more generally, the parameters) of the model in order to tune the model.
However, we believe this possible evidence of Applicant being its own lexicographer is/should be outweighed by the factors mentioned above, namely, (1) Applicant's apparent intent to be in conformity with the general knowledge/ understanding of the art, as reflected in Applicant's disclosure, (2) the principle of charity, and (3) (flowing from (1) and (2)) the interest of compact prosecution. Accordingly, the claim language "tuning the risk indicator" may be considered to be merely an infelicitous usage.
Therefore, in view of the above discussion and the three factors indicated, we interpret the tuning step as saying that the risk indicator is tuned via the tuning of the model, i.e., via identifying of the optimal hyperparameter values (in other words, optimizing the hyperparameter values), which improves the performance of the model and thereby generates an improved (i.e., tuned) model output (i.e., a tuned risk indicator). On this interpretation of the tuning step as a whole, the word "tuning" as used in the phrase "tuning the risk indicator" is interpreted as per the ordinary meaning of "tune," namely, to adjust, modify, or the like. In short, our interpretation is that the risk indicator is tuned (i.e., adjusted) as a necessary consequence of the tuning of the model, the tuning of the model being effected by optimizing the hyperparameters.
Claim Objections
Claims 22 and 41 are objected to because of the following informalities:
Claims 22 and 41 recite:
training the machine learning model based on additional information, the additional information including at least one of additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and
tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model.
The term "additional" is a relative term (i.e., additional to what?).
In light of specification ¶ 0071, the preceding step of claims 22 and 41, and Applicant's remarks9, it is believed that Applicant intended that the recited "additional information" is additional to the information retained in the preceding claim step ("retaining … information associated with one or more previously generated risk indicators"). For purposes of examination, the recited "additional information" is interpreted as such.
In light of Fig. 14 (note 1402-1407), specification ¶¶ 0072-0080, and the rest of claims 22 and 41, it is believed that Applicant intended that the recited "additional transaction data" is additional to the "processed action" and/or to the "supplemental input" that "includes historical data including a chronological listing of all transactions relating to a current account of the user," recited previously in the claims. For purposes of examination, the recited "additional transaction data" is interpreted as such.
Claim 32 is objected to because of the following informalities:
Claim 32 recites:
train the machine learning model based on additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and
The term "additional" is a relative term (i.e., additional to what?).
In light of Fig. 14 (note 1402-1407), specification ¶¶ 0072-0080, and the rest of claim 32, it is believed that Applicant intended that the recited "additional transaction data" is additional to the "processed action" and/or to the "supplemental input" that "includes historical data including a chronological listing of all transactions relating to a current account of the user," recited previously in the claim. For purposes of examination, the recited "additional transaction data" is interpreted as such.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 22-41 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor, or for pre-AIA the applicant regards as the invention.
Unclear Scope/Lack of Antecedent Basis/Relative Terminology/Unclear Antecedent Basis
Claims 22, 32 and 41 recite:
tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model.
Despite the claim interpretation adopted for purposes of examination (see Claim Interpretation section, above), the tuning step on its face admits of two meanings, namely:
(1) the tuning of the risk indicator serves to identify the optimal hyperparameter values
(2) the risk indicator is tuned via the tuning of the model, i.e., via identifying of the optimal hyperparameter values (in other words, optimizing the hyperparameter values), which improves the performance of the model and thereby generates an improved (i.e., tuned) model output (i.e., a tuned risk indicator). (In short, the risk indicator is tuned (i.e., adjusted) as a necessary consequence of the tuning of the model, the tuning of the model being effected by optimizing the hyperparameters.)
This prima facie ambiguity of the meaning of the tuning step renders the claim meaning/ scope unclear and indefinite.
Claims 23 and 33 recite:
wherein the at least one supplemental input further includes at least one of transactional data or a customer characteristic.
The relationship between the "transactional data" (underlined above) recited in claims 23 and 33 and the "additional transactional data" recited in base claims 22 and 32 is not clear (put in other words, the antecedent basis of the underlined language is not clear). For example, it is not clear if the recited "transactional data" of claims 23 and 33 is the same as, or is part of, the "additional transactional data" recited in base claims 22 and 32. (Consequently, for claims 23 and 33, it is not clear, for example, if the training is based on the recited "transactional data" of claims 23 and 33 or not.)
The relationship between the "customer characteristic" (underlined above) recited in claims 23 and 33 and the "customer characteristics" recited in base claims 22 and 32 is not clear (put in other words, the antecedent basis of the underlined language is not clear). For example, it is not clear if the recited "customer characteristic" of claims 23 and 33 is one of the "customer characteristics" recited in base claims 22 and 32. (Consequently, for claims 23 and 33, it is not clear, for example, if the training is based on the recited "customer characteristic" of claims 23 and 33 or not.)
Accordingly, the meaning/scope of the claims is unclear and indefinite.
Claims 24 and 34 depend from claims 23 and 33, respectively, and recite "the transactional data." It is not clear if this recitation refers to the "transactional data" of claims 23 and 33 or to the "additional transactional data" of claims 22 and 32, or to both. In the previous version of the claims and in the original claims, this recitation referred to the "transactional data" of claims 23 and 33. If this recitation refers to the "transactional data" of claims 23 and 33, then it is indefinite for the same reasons as set forth above with respect to the same recitation in claims 23 and 33.
Accordingly, the meaning/scope of the claims is unclear and indefinite.
Claims 25 and 35 depend from claims 23 and 33, respectively, and recite "the customer characteristic." It is not clear if this recitation refers to the "customer characteristic" of claims 23 and 33 or to one of the "customer characteristics" of claims 22 and 32, or to both. In the previous version of the claims and in the original claims, this recitation referred to the "customer characteristic" of claims 23 and 33. If this recitation refers to the "customer characteristic" of claims 23 and 33, then it is indefinite for the same reasons as set forth above with respect to the same recitation in claims 23 and 33.
Accordingly, the meaning/scope of the claims is unclear and indefinite.
Claims 27 and 37 recite "the historical data." Base claims 22 and 32 recite two different kinds of historical data, namely, "historical data including a chronological listing of all transactions relating to a current account of the user" (in the step of receiving a supplemental input) and "historical data associated with the previously generated risk indicators" (in the training step). It is not clear which of these two kinds of historical data in the base claims the recitation of "the historical data" in claims 27 and 37 refers to. (Consequently, for claims 27 and 37, it is not clear, for example, if the supplemental input, based on which the risk indicator is generated, includes the historical data of claims 27 and 37, or if the training is based on the historical data of claims 27 and 37.)
Accordingly, the meaning/scope of the claims is unclear and indefinite.
Claim 32 recites:
tune the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model.
The term "the additional information" lacks antecedent basis and is a relative term (i.e., additional to what?).
In this regard, it is noted that claim 32 differs from claims 22 and 41 in respect of the second training step, which precedes the above-quoted tuning step. Specifically, claims 22 and 41 recite:
training the machine learning model based on additional information, the additional information including at least one of additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and
while claim 32 recites:
train the machine learning model based on additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and
Thus, claim 32 omits the language "additional information, the additional information including at least one of" (as underlined above for claims 22 and 41) in the second training step, while claims 22 and 41 include this language in the second training step. It is believed that this difference may be an oversight/clerical error and that Applicant may have intended claim 32 to recite the same language as claims 22 and 41 in this regard.
In light of specification ¶ 0071, Applicant's remarks (as discussed above with respect to the claim objections to claims 22 and 41), and the rest of claim 32, it is believed that Applicant may have intended that the recited "additional information" is information additional to the information retained in the retaining step ("retain … information associated with one or more previously generated risk indicators"). For purposes of examination, the recited "additional information" is interpreted as such.
However, the claims as drafted do not on their face indicate what the recited "additional information" is (lack of antecedent basis) or what it is additional to (relative terminology).
In particular, for example, as claim 32 is drafted, the claim is not clear on its face as to whether the recited "additional information" refers to solely the "additional transaction data" (recited in the training step) or rather (as in claims 22 and 41) to the "additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators" (recited in the training step).
Accordingly, "the additional information" lacks a clear referent.
In view of the above considerations as to lack of antecedent basis and unclear relative terminology, the claim is indefinite.
Claims 23-31 and 33-40 are (also) rejected by virtue of their dependency from a rejected claim.
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 22-41 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (i.e., a law of nature, a natural phenomenon, or an abstract idea) without significantly more.
Claims 22-41 are directed to a method, system, or non-transitory computer-readable medium, which are/is one of the statutory categories of invention. (Step 1: YES)
Claims 22, 32 and 41 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The claims recite a method, system, and non-transitory computer-readable medium for determining a likelihood of fraud and preventing fraud.
For claims 22, 32 and 41 (claim 32 being deemed representative), the limitations (indicated below in bold) of:
at least one processor configured to:
receive, by the at least one processor, from a transaction channel, a processed action of a user;
receive, by the at least one processor, from at least one external database, at least one supplemental input associated with at least one of the transaction channel, the processed action, or the user, wherein the at least one supplemental input includes historical data including a chronological listing of all transactions relating to a current account of the user;
generate, by a machine learning model executed by the at least one processor, a risk indicator based on the processed action and the at least one supplemental input;
responsive to a determination that the risk indicator exceeds a predetermined threshold, generate an alert indicating a probability of an unauthorized action, wherein the alert and the processed action are placed into an ordered list;
retrieve the processed action from the ordered list and selectively stop, flag for analyst review, or allow execution of the processed action based on the risk indicator; and
retain, by the machine learning model executed by the at least one processor, information associated with one or more previously generated risk indicators;
training the machine learning model based on additional information, the additional information including at least one of additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and
tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model.
as drafted, constitute a process that, under the broadest reasonable interpretation, covers "certain methods of organizing human activity," specifically, "fundamental economic practices or principles" and/or "commercial or legal interactions," but for recitation of generic computer components and generally linking the use of a judicial exception to a particular technological environment or field of use. The Examiner notes that "fundamental economic practices" or "fundamental economic principles" describe concepts relating to the economy and commerce, including hedging, insurance, and mitigating risks, and "commercial interactions" or "legal interactions" include agreements in the form of contracts, legal obligations, advertising, marketing or sales activities or behaviors, and business relations. MPEP 2106.04(a)(2)II.A.,B. If a claim limitation, under its broadest reasonable interpretation, covers "fundamental economic practices or principles" and/or "commercial or legal interactions," but for recitation of generic computer components and generally linking the use of a judicial exception to a particular technological environment or field of use, then it falls within the "certain methods of organizing human activity" grouping of abstract ideas. Accordingly, claims 22, 32 and 41 recite an abstract idea. (Step 2A - Prong 1: YES. The claims recite an abstract idea.)
This judicial exception is not integrated into a practical application. Claims 22, 32 and 41 recite the additional elements of a computing system including at least one processor (the foregoing recited in claim 22), at least one processor configured to: (perform actions) (the foregoing recited in claim 32), one or more instructions that, when executed by the at least one processor of the computing system, cause the computing system to (perform actions) (the foregoing recited in claim 41), the at least one processor, at least one external database, a machine learning model executed by the at least one processor, training the machine learning model based on additional information, and tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model (the foregoing recited in claims 22, 32 and 41), that implement the abstract idea. These additional elements are not described by the applicant and they are recited at a high level of generality (i.e., one or more generic computer elements performing generic computer functions, or generally linking the use of a judicial exception to a particular technological environment or field of use), such that they amount to no more than mere instructions to apply the exception using generic computer elements (namely, all of the additional elements), or such that they amount to no more than generally linking the use of a judicial exception to a particular technological environment or field of use (namely, a machine learning model executed by the at least one processor, training the machine learning model based on additional information, and tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model). Accordingly, even in combination 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. (Step 2A - prong 2: NO. The additional elements do not integrate the abstract idea into a practical application.)
The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception itself. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of a computing system including at least one processor (the foregoing recited in claim 22), at least one processor configured to: (perform actions) (the foregoing recited in claim 32), one or more instructions that, when executed by the at least one processor of the computing system, cause the computing system to (perform actions) (the foregoing recited in claim 41), the at least one processor, at least one external database, a machine learning model executed by the at least one processor, training the machine learning model based on additional information, and tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model (the foregoing recited in claims 22, 32 and 41), to perform the noted steps amount to no more than mere instructions to apply the exception using generic computer elements or generally linking the use of a judicial exception to a particular technological environment or field of use. Mere instructions to apply an exception using generic computer elements or generally linking the use of a judicial exception to a particular technological environment or field of use cannot provide an inventive concept ("significantly more"). Accordingly, even in combination, these additional elements do not provide significantly more. As such, claims 22, 32 and 41 are not patent eligible. (Step 2B: NO. The claims do not provide significantly more.)
Dependent claims 23-31 and 33-40 are similarly rejected because they further define/narrow the abstract idea of independent claims 22, 32 and 41 as discussed above, and/or do not integrate the abstract idea into a practical application or provide an inventive concept such as would render the claims eligible, whether each is considered individually or as an ordered combination.
As for further defining/narrowing the abstract idea:
Dependent claims 23 and 33 merely further describe wherein the at least one supplemental input further includes at least one of transactional data or a customer characteristic.
Dependent claims 24 and 34 merely further describe wherein the transactional data includes at least one of an amount, an erasure, a handwriting change, a signature mismatch, a deposit type, a deposit time, or a deposit location.
Dependent claims 25 and 35 merely further describe wherein the customer characteristic comprises alternate account information, wherein the alternate account is associated with the user.
Dependent claims 26 and 36 merely further describe wherein the alternate account information includes at least one of an account type, a name, a current balance, an opening date, or an account number.
Dependent claims 27 and 37 merely further describe wherein the processed action of the user corresponds to the current account of the user, and wherein the historical data further comprises past information relating to the current account of the user.
Dependent claims 28 and 38 merely further describe wherein the past information relating to the current account of the user includes an account age, an account balance history, an account overdraft history, and an account statement.
Dependent claims 29 and 39 merely further describe wherein the predetermined threshold is based on at least one of a deposit type, a deposit amount, a deposit location, and a history of fraud.
Dependent claims 30 and 40 merely further describe wherein the risk indicator is further generated based on a log-norm scaling of a profile of the user against the user’s profile.
As for additional elements:
Dependent claim 31 recites "tuning the machine learning model based on the at least one supplemental input." This recitation is at a high level of generality such that it amounts to no more than mere instructions to apply the exception using a generic computer element or such that it amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use. Even in combination these additional elements do not integrate the abstract idea into a practical application and do not amount to significantly more than the abstract idea itself.
Claims 23-30 and 33-40 do not recite any additional elements, and accordingly, for the reasons provided above with respect to the independent claims, are not patent eligible.
Therefore, dependent claims 23-31 and 33-40 are not patent eligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries set forth in Comeaux v. John Deere Co., 383 U.S. 1, 148 USPQ 459 (1966), that are applied 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 22-27, 29, 31-37, 39 and 41 are rejected under 35 U.S.C. 103 as being unpatentable over Comeaux et al. (U.S. Patent No. 11,669,844), hereafter Comeaux, in view of Handelman (U.S. Patent Application Publication No. 2021/0326881 A1), further in view of Abifaker et al. (U.S. Patent Application Publication No. 2015/0278817 A1), hereafter Abifaker, and further in view of Shevyrev et al. (U.S. Patent Application Publication No. 2023/0281629 A1), hereafter Shevyrev.
Regarding Claims 22, 32 and 41
Comeaux teaches:
(claim 22) the method being performed by at least one processor (claim 1: steps of the method are performed "by a/the computing system"; claim 11: "one or more processors; and one or more memories storing instructions that, when executed by the one or more processors, cause the computing system to perform a process comprising")
(claim 32) at least one processor configured to: (5:53-57; claim 11)
(claim 41) A non-transitory computer-readable medium storing a set of instructions for identifying an unauthorized action in a computing system including at least one processor, the set of instructions comprising: one or more instructions that, when executed by the at least one processor of the computing system, cause the computing system to: (5:53-57, 20:5-35, claim 6)
receiving, by the at least one processor (5:53-54 processor of alert-generating server 102), from a transaction channel (17:10-13), a processed action of a user; (11:14-16, 11:26-27, 12:6-27; Fig. 2, 204, 16:40-42)
receive, by the at least one processor, from at least one external database, at least one supplemental input associated with at least one of the transaction channel, the processed action, or the user; (9:29-10:20 behavior profile of user, and/or the data it contains, namely, user personal data, user financial data, and user social network data, teaches at least one supplemental input associated with at least one of the transaction channel, the processed action, or the user; regarding from at least one external database: 16:11-19 "To generate the behavior profile of the user, the alert-generating server or any other server of the system may obtain personal data, financial data, and social network data from internal and external databases of the system."; per 9:29-31, 12:59-13:7, 15:5-9, behavior profile of user may be stored in database 104, and per 12:59-13:7 the webserver (12:43-58, i.e., system server 106) may access user data in the behavior profile in the database, which teaches receiving (accessing), from the external database, the supplemental input (behavior profile); in addition further regarding external: 6:39-41, 13:36-39, 14:46-60 database 104 (which stores behavior profile) may be hosted on any server in system 100, and may be separate from alert generating server 102, therefore under broadest reasonable interpretation database 104 may be an "external" database, e.g., relative to alert generating server 102)
wherein the at least one supplemental input includes historical data …; (10:62-11:1 , 11:30-32)
generate, by a machine learning model (4:49-52 teaches that the alert generation model is a machine-learning model) executed by the at least one processor, a risk indicator (alert probability score) based on the processed action and the at least one supplemental input; and (2:43-49, 11:26-32, Fig. 2, 210, 17:60-18:14; regarding based on … the at least one supplemental input: 16:1-3 "The alert-generating server may generate the alert-generation model using one or more inputs from a behavior profile of each user of a system (the at least one supplemental input)." -- since the alert probability score (risk indicator) is generated by the alert generation model, and the alert generation model is generated using behavior profile inputs, it follows that the alert probability score (risk indicator) is generated based on the behavior profile (the at least one supplemental input))
responsive to a determination that the risk indicator exceeds a predetermined threshold, generate an alert indicating a probability of an unauthorized action, (2:49-53, 11:32-35, Fig. 2, 212, 18:21-18:50)
wherein the alert and the processed action are placed into an ordered list; (regarding wherein the alert and the processed action are placed into: 14:46-50, "Database 104 … is capable of storing the fraudulent events [processed action] and alert"; 14:59-63; regarding an ordered list: 4:65-67 "The alert-generating systems and apparatuses may sort the fraudulent events according to a priority [an ordered list], based on the relative fraud probability scores."; 5:8-11 "Analyst computers may query and fetch these alerts from a database … according to a priority level [an ordered list] of the respective alerts."; 14:59-63 "The database 104 may … include a non-transitory machine-readable storage media capable of receiving and storing fraudulent events and alerts. The database 104 may have a logical construct of data files that are stored in non-transitory machine-readable storage media, … for various data queries …." -- the fraudulent events and alerts may be queried from the database according to a priority [an ordered list according to which they were placed in the database])
retrieving the processed action from the ordered list and selectively stopping, … or allowing execution of the processed action based on the risk indicator; and (11:32-35; 18:21-61)
retaining, by the machine learning model executed by the at least one processor, information associated with one or more previously generated risk indicators; (9:19-22 "Therefore, the initial alert-generation model may need to be trained using other known fraudulent users and/or activities to achieve optimum accuracy." – i.e., this data is retained for training/tuning)
(claims 22 and 41) training the machine learning model based on additional information, the additional information including at least one of additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and (2:60-62 "the server trains the alert-generation model using the updated behavior profile"; 9:35-37 the "behavior profile of the user may include, but not limited to, user personal data, user financial data, and user social network data." – these various data are further described at 9:38-10:20, e.g., 9:60-63 "the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud.")
(claim 32) train the machine learning model based on additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and (2:60-62 "the server trains the alert-generation model using the updated behavior profile"; 9:35-37 the "behavior profile of the user may include, but not limited to, user personal data, user financial data, and user social network data." – these various data are further described at 9:38-10:20, e.g., 9:60-63 "the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud.")
…
As indicated above, Comeaux teaches a portion of the following limitation but does not explicitly disclose it in its entirety. However, Handelman teaches:
… wherein the at least one supplemental input includes historical data including a chronological listing of all transactions relating to a current account of the user; (claim 3, 0016, 0045, 0057; for context see abstract, 0002, 0011-0023, 0045-0056, Fig. 3, 0057-0069, Fig. 4)
It would have been obvious to one of ordinary skill in the art not later than the effective filing date of the claimed invention to have modified Comeaux's systems and methods for detecting and preventing fraudulent events, by incorporating therein these teachings of Handelman regarding using a chronological listing of all transactions as a factor for fraud detection, because this data of Handelman is known to be a relevant factor in fraud detection and therefore adding it to/including it in the set of input data Comeaux already uses would render Comeaux's input data more comprehensive/ robust, thus tending to yield more accurate outputs, and accordingly to improve Comeaux's performance, see Handelman, e.g., 0002, 0011-0023.
As indicated above, Comeaux teaches a portion of the following limitation but Comeaux in view of Handelman does not explicitly disclose it in its entirety. However, Abifaker teaches:
… selectively stopping (e.g., 0068; Fig. 2, 230; Fig. 3, "Fail"), flagging for analyst review (e.g., 0070 "determine that manual review … is needed"; Fig. 2, 245; Fig. 3, "Human Audit"), or allowing execution of (e.g., 0069; Fig. 2, 250; Fig. 3, "Pass") the processed action based on the risk indicator; and (0067-0070, 0075-0076, Fig. 2, 225, 230, 245, 250, Fig. 3, Fig. 5, 540)
It would have been obvious to one of ordinary skill in the art not later than the effective filing date of the claimed invention to have modified the combination of Comeaux's systems and methods for detecting and preventing fraudulent events, as modified by Handelman's teachings regarding using a chronological listing of all transactions as a factor for fraud detection, by incorporating therein these teachings of Abifaker regarding a tripartite scheme (stop/cease, flag for review, allow/proceed) for handling a transaction depending upon a fraud risk score, because it provides a more comprehensive, fine-tuned risk assessment/scoring and consequent responsive actions (as compared to Comeaux's bipartite scheme), so as to more likely treat transactions appropriately according to their actual risk level, thus leading to more effective and satisfactory outcomes (e.g., medium risk transactions would be more likely to receive an intermediate level of responsive action, namely, a review, rather than being approved outright as if they were low risk or rejected outright as if they were high risk), see Comeaux and Abifaker, citations provided for the claim limitation in question.
As indicated above, Comeaux teaches the retaining step and the training step. In addition, Comeaux teaches tuning the alert generation model (neural network) (4:49-64), but Comeaux in view of Handelman and Abifaker does not explicitly disclose the tuning step in its entirety. However, Shevyrev remedies this deficiency, teaching all three of these steps (retaining, training and tuning steps):
retaining, by the machine learning model (Fig. 4, 308; 0028, 0078) executed by the at least one processor (Fig. 8, 802; 0106), information associated with one or more previously generated risk indicators to tune the risk indicator; (0072 "the mobile check deposit system 102 either denies the network transaction at an act 318 or approves the network transaction at an act 320. For example, at the act 318, the mobile check deposit system 102 denies the network transaction by changing the temporary suspension of the network transaction to a rejection. For example, the mobile check deposit system 102 labels the network transaction as fraudulent (or rejected check) and rejects the network transaction from issuing or completing. In one or more embodiments, the mobile check deposit system 102 saves [retains] the rejected network transaction and corresponding data for training purposes (e.g., as described below in relation to FIG. 4 )."; see also 0073 "At the act 320, the mobile check deposit system 102 approves the network transaction in response to successful verification processes by unsuspending the network transaction. In one or more embodiments, unsuspending the network transaction allows the network transaction to issue or complete (e.g., such that funds between network accounts settle). … Additionally, in one or more embodiments, the mobile check deposit system 102 whitelists [hence, retains] one or more network accounts and/or similar transactions associated with a network account (e.g., to reduce or prevent future false positives)."; see also 0080-0081)
(claims 22 and 41) training the machine learning model based on additional information, the additional information including at least one of additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and (0076-0081 with reference to Fig. 4: training of the machine learning model is performed using training features, as enumerated in 0077 and Table 1, where the enumerated features include the recited additional information: customer characteristics is taught by, e.g., customer-service-contact data (0077), risk score based on phone number/phone interactions/email address/email correspondence/address, # of disputes, correlation score re name and phone number/address (Table 1), and additional transactional data and historical data associated with the previously generated risk indicators is taught by many of the enumerated features)
(claim 32) train the machine learning model based on additional transactional data, customer characteristics, and historical data associated with the previously generated risk indicators; and (0076-0081 with reference to Fig. 4: training of the machine learning model is performed using training features, as enumerated in 0077 and Table 1, where the enumerated features include the recited additional information: customer characteristics is taught by, e.g., customer-service-contact data (0077), risk score based on phone number/phone interactions/email address/email correspondence/address, # of disputes, correlation score re name and phone number/address (Table 1), and additional transactional data and historical data associated with the previously generated risk indicators is taught by many of the enumerated features)
tuning the risk indicator, using the retained information and the additional information, wherein the tuning occurs based on the training and identifies optimal values of one or more hyperparameters to maximize performance of the machine learning model. (0080-0081, Fig. 4, specifically: regarding the tuning … identifies optimal values of one or more hyperparameters: 0081 "the mobile check deposit system 102 adjusts various parameters/ hyperparameters"; regarding to maximize performance of the machine learning model: 0081 "In so doing, the mobile check deposit system 102 can improve the quality/accuracy of training check-return predictions in subsequent training iterations—by narrowing the difference between training check-return predictions and ground truth check data in subsequent training iterations."; regarding based on the training: 0080-0081, Fig 4: the adjusting of the hyperparameters is "based upon" the "losses 410" (or the returning of the losses 410 to the model 308 by the loss function 408) (0081), the losses 410 having been determined by the loss function 408 based on comparison of training check-return predictions 405 with ground truth check data 406, which is performed during training (0080), hence the adjusting of the hyperparameters is based (by the foregoing chain of operations/chain of reasoning) on the training; regarding tuning the risk indicator: 0081 "In so doing, the mobile check deposit system 102 can improve the quality/accuracy of training check-return predictions [risk indicator] in subsequent training iterations—by narrowing the difference between training check-return predictions and ground truth check data in subsequent training iterations." – i.e., the output of the machine learning model, namely, the check-return prediction (risk indicator), is improved in quality/accuracy (i.e., tuned), by virtue of the tuning of the model achieved (the tuning of the model being effected by adjusting the hyperparameters); regarding using the retained information: as explained above (see "regarding based on the training"), the tuning is based on the training, and, as also explained above (see the retaining step), as per 0072, 0073, outcome information and associated transaction information (the retained information) is used for training, therefore it follows that the tuning is based on and, under broadest reasonable interpretation, uses this outcome information and associated transaction information (the retained information); regarding using … the additional information: as explained above (see "regarding based on the training"), the tuning is based on the training, and, as also explained above (see the training step), as per 0077 and Table 1, the training uses the additional information, therefore it follows that the tuning is based on and, under broadest reasonable interpretation, uses the additional information)
It would have been obvious to one of ordinary skill in the art not later than the effective filing date of the claimed invention to have modified the combination of Comeaux's systems and methods for detecting and preventing fraudulent events, as modified by Handelman's teachings regarding using a chronological listing of all transactions as a factor for fraud detection, and as further modified by Abifaker's teachings regarding a tripartite scheme (stop/cease, flag for review, allow/proceed) for handling a transaction depending upon a fraud risk score, by incorporating therein these teachings of Shevyrev regarding training and tuning (adjusting hyperparameters of a machine learning model to improve performance) based on retained previous outputs and other data,10 because it would improve performance of Comeaux's machine learning model in a way that goes beyond Comeaux's training/tuning of the machine learning model, see Comeaux and Shevyrev, citations provided for the claim limitation in question, MPEP 2143.I.C.,D.
Regarding Claims 23 and 33
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 as set forth above. Comeaux further teaches:
wherein the at least one supplemental input further includes at least one of transactional data or a customer characteristic. (9:33-10:20)
Regarding Claims 24 and 34
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 and intervening claims 23 and 33 as set forth above. Comeaux further teaches:
wherein the transactional data includes at least one of an amount, an erasure, a handwriting change, a signature mismatch, a deposit type, a deposit time, or a deposit location. (9:62 account balance teaches amount)
Regarding Claims 25 and 35
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 and intervening claims 23 and 33 as set forth above. Comeaux further teaches:
wherein the customer characteristic comprises alternate account information, wherein the alternate account is associated with the user. (9:58-63 "the alert-generating server 102 generate [sic] a behavior profile of the user based on at least the personal data and also the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud." – "types of accounts" (plural) indicates the user has multiple accounts; as such, one of the multiple accounts is the account involved in the current transaction (the "current account" recited in claims 22 and 32) (see claim 27), and another of the multiple accounts is an alternate account; 13:17-20 reconfirms that the user has multiple accounts ("For instance, a user may open multiple new accounts associated with a user identifier, or the user may move unusually large amounts of money between accounts or to accounts at a different financial institution."); 16:19-24 reconfirms that the behavior profile may include events such as opening an account and, per 9:58-63, 13:17-20, there may be multiple of such events, hence one such event pertains to an alternate account)
Regarding Claims 26 and 36
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 and intervening claims 23, 25, 33 and 35 as set forth above. Comeaux further teaches:
wherein the alternate account information includes at least one of an account type, a name, a current balance, an opening date, or an account number. (9:58-63 "the alert-generating server 102 generate [sic] a behavior profile of the user based on at least the personal data and also the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud.")
Regarding Claims 27 and 37
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 and intervening claim 23 as set forth above. Comeaux further teaches:
wherein the processed action of the user corresponds to the current account of the user, (Fig. 2, 204, 16:40-42 fraud event (the processed action of the user) pertains to user's current account, as evidenced by the language "account takeover," "account identifiers," etc. (corresponds to a current account of the user))
and wherein the historical data further comprises past information relating to the current account of the user. (2:29-31 "the behavior profile of the user comprises at least a record of events previously undertaken by the user in an account of the user."; 16:19-24 "The alert-generating server … may use a record of events previously undertaken by the user in an account of the user to generate a behavior profile of each user. The events may include, but not limited to, transfer funds, open a new account, withdraw funds, check funds, etc. in his/her account."; 9:58-63 "In some embodiments, the alert-generating server 102 generate [sic] a behavior profile of the user based on at least the personal data and also the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud.")
Regarding Claims 29 and 39
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 as set forth above. Comeaux further teaches:
wherein the predetermined threshold is based on at least one of a deposit type, a deposit amount, a deposit location, and a history of fraud. (18:27-32 "In some embodiments, a pre-defined threshold score may be based on a profile of a user on whose account the fraud event has been detected. In some embodiments, a pre-defined threshold score may be based on one or more attributes of fraud event such as illegal fund transfer request or probable unauthorized access of the account."; 9:58-63 "In some embodiments, the alert-generating server 102 generate [sic] a behavior profile of the user based on at least the personal data and also the financial data of the user that may include, but not limited to, types of accounts, account balance, credit score, any known flag for activity or known fraud." -- threshold is based on profile, which in turn is based on history of fraud, so threshold is based on history of fraud)
Regarding Claim 31
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claim 22 as set forth above. Comeaux further teaches:
tuning the machine learning model based on the at least one supplemental input. (4:49-64 e.g., "This allows the alert generation model to automatically validate new events and tune itself accordingly." – the validation of "new" events indicates that the model has received new data (supplemental input) beyond an initial transaction; although they do not use the word "tune," the following passages teach similar substantive content: 8:39-62 reconfigure model based on new data (input); 19:12-35 updating and reconfiguring the model based on updated behavior profile (input)11)
Claims 28 and 38 are rejected under 35 U.S.C. 103 as being unpatentable over Comeaux et al. (U.S. Patent No. 11,669,844), hereafter Comeaux, in view of Handelman (U.S. Patent Application Publication No. 2021/0326881 A1), further in view of Abifaker et al. (U.S. Patent Application Publication No. 2015/0278817 A1), hereafter Abifaker, further in view of Shevyrev et al. (U.S. Patent Application Publication No. 2023/0281629 A1), hereafter Shevyrev, and further in view of Carson et al. (U.S. Patent Application Publication No. 2022/0138844 A1), hereafter Carson.
Regarding Claims 28 and 38
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 and intervening claims 23, 27 and 37 as set forth above. Comeaux further teaches:
wherein the past information relating to the current account of the user includes an account age, an account balance history, … and an account statement. (9:58-63 "In some embodiments, the alert-generating server 102 generate [sic] a behavior profile of the user based on at least the personal data and also the financial data of the user that may include, but not limited to, types of accounts, account balance [account balance history], credit score, any known flag for activity or known fraud."; 16:11-24 "The alert-generating server or any other server of the system may use a record of events previously undertaken by the user in an account of the user to generate a behavior profile of each user. The events may include, but not limited to, transfer funds, open a new account, withdraw funds, check funds, etc. in his/her account."; opening an account is also taught by 13:17-18 -- an event in the record of events comprising the opening of a new account indicates the account age; under broadest reasonable interpretation, the "record of events previously undertaken by the user in an account of the user" teaches an account statement)
Comeaux in view of Handelman, Abifaker, and Shevyrev does not explicitly disclose but Carson teaches:
wherein the past information relating to the current account of the user includes an account age (e.g., 0012), an account balance history (e.g., 0012, 0015), an account overdraft history (e.g., 0022), …. (Abstract, 0007, 0012, 0017, 0022, claim 8, performing fraud review/determining legitimacy of transaction includes checking overdraft history (an account overdraft history) as factor for determining likelihood of fraud) (Note Carson's teachings of "an account age" and "an account balance history" are cited as alternative teachings to those of Comeaux cited above.)
It would have been obvious to one of ordinary skill in the art not later than the effective filing date of the claimed invention to have modified the combination of Comeaux's systems and methods for detecting and preventing fraudulent events, as modified by Handelman's teachings regarding using a chronological listing of all transactions as a factor for fraud detection, as further modified by Abifaker's teachings regarding a tripartite scheme (stop/cease, flag for review, allow/proceed) for handling a transaction depending upon a fraud risk score, and as further modified by Shevyrev's teachings regarding training and tuning (adjusting hyperparameters of a machine learning model to improve performance) based on retained previous outputs and other data, by incorporating therein these teachings of Carson regarding checking overdraft history as a factor for determining likelihood of fraud / for determining legitimacy of a transaction, because overdraft history is known to be a relevant factor in fraud detection and therefore adding it to/including it in the set of input data Comeaux already uses would render Comeaux's input data more comprehensive/robust, thus tending to yield more accurate outputs, and accordingly to improve Comeaux's performance, see Carson, citations given.
Claims 30 and 40 are rejected under 35 U.S.C. 103 as being unpatentable over Comeaux et al. (U.S. Patent No. 11,669,844), hereafter Comeaux, in view of Handelman (U.S. Patent Application Publication No. 2021/0326881 A1), further in view of Abifaker et al. (U.S. Patent Application Publication No. 2015/0278817 A1), hereafter Abifaker, further in view of Shevyrev et al. (U.S. Patent Application Publication No. 2023/0281629 A1), hereafter Shevyrev, and further in view of Pavlovic ("Log-normal Distribution - A simple explanation”).
Regarding Claims 30 and 40
Comeaux in view of Handelman, Abifaker, and Shevyrev teaches the limitations of base claims 22 and 32 as set forth above. Comeaux further teaches
wherein the risk indicator (alert probability score) is further generated based on … a profile of the user (behavior profile of a user) … the user's profile. (2:26-62 generates alert probability score (2:44) based on alert generation model (2:46) which is based on a behavior profile of a user (2:26-31); 9:29-1020 describes the behavior profile)
Comeaux in view of Handelman, Abifaker, and Shevyrev does not explicitly disclose but Pavlovic teaches:
… a log-norm scaling of … against …. (Note Applicant’s specification (¶ 0045) defines “log-norm scaling” thus: “Log-norm scaling may refer to applying a logarithmic transformation to values, which transforms the values onto a scale that approximates the normality”; Pavlovic, p. 3, teaches the same thing: “Let’s say your data [values] fits a log-normal distribution. If you then take the logarithm of [applying a logarithmic transformation to] all your data points [values], the newly transformed [logarithmically transformed] points [values] will now fit a normal distribution [a scale that approximates the normality].”)
It would have been obvious to one of ordinary skill in the art not later than the effective filing date of the claimed invention to have modified the combination of Comeaux's systems and methods for detecting and preventing fraudulent events, as modified by Handelman's teachings regarding using a chronological listing of all transactions as a factor for fraud detection, as further modified by Abifaker's teachings regarding a tripartite scheme (stop/cease, flag for review, allow/proceed) for handling a transaction depending upon a fraud risk score, and as further modified by Shevyrev's teachings regarding training and tuning (adjusting hyperparameters of a machine learning model to improve performance) based on retained previous outputs and other data, by incorporating therein these teachings of Pavlovic regarding log-norm scaling, because these teachings of Pavlovic are a known way to model various natural phenomena, see Pavlovic, p. 1, and they have particular advantages under certain circumstances applicable to Comeaux, e.g., where the data cannot be negative (e.g., data such as a probability of fraud / a risk value), where the data skews positive, with most values clustered near the low end and a long tail extending rightward to occasional high outliers (e.g., transaction data, which comprises mostly non-fraudulent transactions and a relatively small amount of fraudulent transactions), and where the data grows multiplicatively/cumulatively (e.g., a plurality of historical transaction data) -- by transforming such data with a logarithm, it can be normalized, allowing for easier analysis with powerful standard statistical techniques, such as calculating z-scores, performing linear regression, estimating parameters using Maximum Likelihood Estimation (MLE), etc.
Conclusion
The prior art made of record and not relied upon, as set forth in the accompanying Notice of References Cited (PTO-892), is considered pertinent to applicant's disclosure.
Comeaux (US-10567402-B1) and Comeaux (US-11722502-B1) teach fraud detection/prevention similar to Comeaux (US-11669844-B1) but to greater depth in certain aspects;
Phatak (US-20220006899-A1) and Anderson (US-12136096-B1) teach a fraud alert queue that prioritizes fraud alerts based on fraud importance;
Vaswani (US-20220377090-A1) teaches fraud detection/prevention (including risk scores and alerts) similar to Comeaux (US-11669844-B1);
Karpovsky (US-20220191173-A1) teaches determining fraud risk based on VPN and/or proprietary knowledge and periodic monitoring; and
Pavlovic ("Log-normal Distribution - A simple explanation”) teaches content about log-normal distribution similar to that of Applicant's disclosure (specification paragraph 0045)
Vimal (US-20230186311-A1) (qualifying as prior art based on Indian priority date) teaches inter alia benchmarking a machine learning model based on precision, recall, F1, and/or F2 scores, see 0097.
Thomas (US-10997596-B1) teaches appending a fraud accuracy tag to a declined transaction, where the fraud accuracy tag is indicative of whether the decline of the transaction is a true positive decline or a false positive decline, whereby the fraud accuracy tag is suitable to provide insight into accuracy of a fraud strategy implemented in connection with the declined transaction.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DOUGLAS W PINSKY whose telephone number is (571)272-4131. The examiner can normally be reached on 8:30 am - 5:30 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, Jessica Lemieux can be reached on 571-270-3445. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/DOUGLAS W PINSKY/
Examiner, Art Unit 3626
1 See the rejection in the body of the Office Action hereinbelow for the precise claim language that is not part of the abstract idea (non-bolded portions).
2 Note the recited "risk indicator" is the output of the machine learning model. That this is so is shown, e.g., by the "generating" step ("generate, by a machine learning model executed by the at least one processor, a risk indicator based on the processed action and the at least one supplemental input"). Relatedly, see the disclosure, ¶ 0081, Fig. 15, showing inputs 1508, model 1509, and risk indicators 1510, which are outputted from model 1509.
3 Note the "additional information" is defined in the training step as being capable of being "information associated with the previously generated risk indicators" (namely, "historical data associated with the previously generated risk indicators"). Therefore, the "additional information" is not necessarily distinct from any and all "information associated with the previously generated risk indicators" (recited in the retaining step). Therefore, it is not necessary to identify all the "information associated with the previously generated risk indicators" (of the retaining step) prior to identifying the "additional information" (of the training step). Therefore, it is not necessary to perform the retaining step prior to the training step.
4 See ¶ 0081, Fig. 15, inputs 1501-1505, 1508, machine learning model 1509, output (risk indicator) 1510.
5 For the content of this paragraph regarding tuning a machine learning model, see, e.g., Khavronin (US-20220188700-A1), 0026-0030; Gorodetski ("Hyperparameter Tuning Methods - Grid, Random or Bayesian Search?"), pp. 3-5.
6 As noted, the recited "risk indicator" is the output of the machine learning model. That this is so is shown, e.g., by the "generating" step ("generate, by a machine learning model executed by the at least one processor, a risk indicator based on the processed action and the at least one supplemental input"). Relatedly, see the disclosure, ¶ 0081, Fig. 15, showing inputs 1508, model 1509, and risk indicators 1510, which are outputted from model 1509.
7 Parker ("Unraveling the Enigmatic Mystery of “The Principle of Charity"), e.g., pp. 1-2.
8 As noted above, the recited risk indicator is the output of the machine learning model.
9 See Applicant's instant Response, p. 17, last paragraph – p. 18, first paragraph, e.g., the argument regarding the chronological sequence of operations -- that the training is based on additional information received after the retention of the information associated with the previously generated risk indicators.
10 Note: various combinations of Shevyrev with Comeaux as modified by the other references are possible. For example, all of the limitations taught by Shevyrev may be incorporated into Comeaux (as modified by the other references), i.e., substituting Shevyrev's teachings of the retaining and training steps for Comeaux's teachings of the retaining and training steps. As an alternative example, Shevyrev's tuning step may be added to Comeaux (as modified by the other references).
11 Note as per the literature the term "reconfigure" is a synonym for "tune" in the context of tuning a machine learning model / tuning parameters/hyperparameters of a machine learning model, see, e.g., Weber (DE 102024111093 A1) p. 7 ("Rather than incurring the significant effort and expense of developing a custom model, the process of hyperparameter tuning offers the opportunity to reconfigure a base model to more closely match certain expectations."); Hayes (US-11301781-B2) 12:42-51 ("The plurality of disparate, optimization worker machines may include any number and type of worker machine that functions to optimize and/or tune one or more aspects of the optimization work request and components (e.g., tuning sources) of the system implementing the method 200. For instance, the plurality of disparate, optimization work machines may include a first optimization worker machine that functions to reconfigure and optimize hyperparameters of member optimization models of an ensemble of optimization models, …."); Pengetnze (US-20190122770-A1) 0021 "The artificial intelligence model tuning process 86 may utilize adaptive self-learning capabilities using machine learning technologies. The capacity for automatic self-reconfiguration enables the system and method 10 to be sufficiently flexible and adaptable to detect and incorporate trends or differences in the underlying patient population that may affect the predictive accuracy of a given algorithm. … The artificial intelligence model tuning process 86 may automatically modify or improve a predictive model in three exemplary ways. … Third, the artificial intelligence model tuning process 86 may, automatically and without human supervision, evaluate new variables present in the data feed but not currently used in the predictive model, which may result in improved accuracy. The artificial intelligence model tuning process 86 may compare the actual observed outcome of the event to the predicted outcome then separately analyze the variables within the model that contributed to the incorrect outcome. It may then re-weigh the variables that contributed to this incorrect outcome, so that in the next reiteration those variables are less likely to contribute to a false prediction. In this manner, the artificial intelligence model tuning process 86 is adapted to reconfigure or adjust the predictive model based on the specific clinical setting or population in which it is applied without manual reconfiguration or modification of the predictive model.").
