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 .
Status of Claims
This action is in reply to the communications filed on 07/23/2024.
Claims 1-20 are currently pending and have been examined.
This action is made Non-Final.
Examiner Request
The Applicant is requested to indicate where in the specification there is support for future claim amendments to avoid U.S.C 112(a) issues that can arise. The Examiner thanks the Applicant in advance.
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 of generating a loan price for a loan applicant, without significantly more.
Examiner has identified claim 10 as the claim that represents the claimed invention presented in independent claims 1, 10, and 19.
Claim 1 is directed to a method, which is one of the statutory categories of invention; Claim 10 is directed to an apparatus, which is one of the statutory categories of invention; and Claim 19 is directed to a non-transitory computer readable storage medium which is one of the statutory categories of invention. (Step 1: YES).
Claim 10 is directed to a computing apparatus for mitigating disparities between outputs of different artificial intelligence / machine learning (AI/ML) models, the computing apparatus comprising: a processor; a memory; and a communication interface coupled to each of the processor and the memory, wherein the processor is configured to: receive, via the communication interface, a first uncertainty value and a first set of model quality-related parameter values that are associated with a first model that is configured to generate a first loan price for a loan applicant; receive, via the communication interface, a second uncertainty value and a second set of model quality-related parameter values that are associated with a second model that is configured to generate a second loan price for the loan applicant; receive, via the communication interface, a set of feature weight functions that relate to weights of target metrics; calculate, for the first model based on the first uncertainty value, the first set of model quality-related parameter values, and the set of feature weight functions, a set of first model weights; calculate, for the second model based on the second uncertainty value, the second set of model quality-related parameter values, and the set of feature weight functions, a set of second model weights; select a customized model based on the set of first model weights and the set of second model weights; receive, via the communication interface, a first tabular set of personal data that comprises a first data subset that relates to individualized financial information associated with each respective applicant from among a plurality of loan applicants and a second data subset that relates to individualized demographic information associated with each respective applicant from among the plurality of loan applicants; train the customized model by using the first tabular set of personal data, the target metrics, a predetermined set of sensitive labels, and historical information that relates to outputs generated by at least one from among the first model, the second model, and the customized model; calculate an updated uncertainty value and an updated set of model quality-related parameter values for the trained customized model; and use the trained customized model to generate a customized loan price for a first loan applicant from among the plurality of loan applicants. These series of steps describe the abstract idea of generating a loan price for a loan applicant (with the exception of the italicized and bolded terms above), which is mitigating the risk of outputting disparities while processing and generating a loan price for a loan applicant among a plurality of loan applicants; therefore, corresponding to a fundamental economic principle or practice (including mitigating risk). Hence, a fundamental economic principle or practice (mitigating risk) is a Certain Methods of Organizing Human Activity. The abstract idea is also the processing of a loan for a loan applicant, which is a commercial interaction. Therefore, a commercial interaction is also a Certain Methods of Organizing Human Activity. The system limitations, e.g., a computing apparatus, artificial intelligence / machine learning (AI/ML) models, processor, memory, communication interface, first model, second model, customized model, and trained customized model do not necessarily restrict the claim from reciting an abstract idea. Thus, claim 10 is directed to an abstract idea (Step 2A-Prong 1: YES).
This judicial exception is not integrated into a practical application because the additional elements of a computing apparatus, artificial intelligence / machine learning (AI/ML) models, processor, memory, communication interface, first model, second model, customized model, and trained customized model are no more than simply applying the abstract idea using generic computer elements. The additional elements listed above are all recited at a high level of generality and under their broadest reasonable interpretation comprises a generic computing arrangement. The presence of a generic computer arrangement is nothing more than to implement the claimed invention (MPEP 2106.05(f)). Therefore, the recitations of additional elements do not meaningfully apply the abstract idea, and hence do not integrate the abstract idea into a practical application. Thus, claim 10 is directed to an abstract idea (Step 2A-Prong 2: NO).
Claim 10 does not include additional elements that are sufficient to amount to significantly more than the judicial exception because the additional elements, a computing apparatus, artificial intelligence / machine learning (AI/ML) models, processor, memory, communication interface, first model, second model, customized model, and trained customized model are recited at a high level of generality in that it results in no more than simply applying the abstract idea using generic computer elements. The additional elements when considered separately and as an ordered combination do not amount to add significantly more as these limitations provide nothing more than to simply apply the exception in a generic computer environment (Step 2B: NO). Thus, claim 10 is not patent eligible.
Similar arguments can be extended to the other independent claims, claims 1 and 19; and hence claims 1 and 19 are rejected on similar grounds as claim 10.
Dependent claims 2-9, 11-18, and 20 are directed to a method, an apparatus, and a non-transitory computer readable storage medium, respectively, which perform the steps that describe the abstract idea of generating a loan price for a loan applicant. Furthermore, dependent claims 9 and 18 are directed to a method and an apparatus, respectively, which recite the steps: receive, via the communication interface, a third uncertainty value and a third set of model quality-related parameter values that are associated with a third model that is configured to generate a third loan price for the loan applicant; and calculate, for the third model based on the third uncertainty value, the third set of model quality-related parameter values, and the set of feature weight functions, a set of third model weights, wherein the selection of the customized model is further based on the set of third model weights, and wherein the training of the customized model is performed by using additional historical information that relates to outputs generated by the third model. These series of steps describe the abstract idea of generating a loan price for a loan applicant, (with the exception of the italicized and bolded terms above), which is mitigating the risk of outputting disparities while processing and generating a loan price for a loan applicant among a plurality of loan applicants; therefore, corresponding to a fundamental economic principle or practice (including mitigating risk). Hence, a fundamental economic principle or practice (mitigating risk) is a Certain Methods of Organizing Human Activity. The abstract idea is also the processing of a loan for a loan applicant, which is a commercial interaction. Therefore, a commercial interaction is also a Certain Methods of Organizing Human Activity. Thus, dependent claims 2-9, 11-18, and 20 are directed to an abstract idea. The additional limitations of a computing apparatus, artificial intelligence / machine learning (AI/ML) models, processor, memory, communication interface, first model, second model, customized model, trained customized model, and third model are no more than simply applying the abstract idea using generic computer elements. Therefore, the recitations of additional elements do not meaningfully apply the abstract idea, and hence, do not integrate the abstract idea into a practical application. Furthermore, the additional elements: a computing apparatus, artificial intelligence / machine learning (AI/ML) models, processor, memory, communication interface, first model, second model, customized model, trained customized model, and third model, do not amount to add significantly more as these limitations provide nothing more than to simply apply the exception in a generic computer environment.
Dependent claims 2-9, 11-18, and 20 have further defined the abstract idea that is present in their respective independent claims 1, 10, and 19; and thus correspond to Certain Methods of Organizing Human Activity, and hence are abstract in nature for the reason presented above. The dependent claims 2-9, 11-18, and 20 do not include any additional elements that integrate the abstract idea into a practical application or are sufficient to amount to significantly more than the judicial exception when considered both individually and as an ordered combination. Therefore, claims 2-9, 11-18, and 20 are directed to an abstract idea.
Thus, claims 1-20 are not patent-eligible.
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 set forth in Graham 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 1-20 rejected under 35 U.S.C. 103 as being unpatentable over Ward (U.S. Patent Application Publication No. US 2024/0046349 A1 hereinafter “Ward”), in view of Wu (U.S. Patent Application Publication No. US 2023/0214695 A1; hereinafter “Wu”).
Regarding Claims 1, 10, and 19:
Ward teaches:
A method for mitigating disparities between outputs of different artificial intelligence / machine learning (AI/ML) models, the method being implemented by at least one processor, the method comprising: (Ward, obtaining output from a machine learning (ML) model responsive to input data, obtaining initial training data representing training data used to train the ML model; fair includes having less disparate impact on protected versus unprotected groups. In some implementations, more fair includes improved accuracy by compensating for training data that under represents one or more constituent groups or elements of the training data. (See, Abstract; Para. 4-5, 91, 97; Fig. 1, 5));
A computing apparatus for mitigating disparities between outputs of different artificial intelligence / machine learning (AI/ML) models, the computing apparatus comprising: a processor; a memory; and a communication interface coupled to each of the processor and the memory, wherein the processor is configured to: (Ward, obtaining output from a machine learning (ML) model responsive to input data, obtaining initial training data representing training data used to train the ML model; fair includes having less disparate impact on protected versus unprotected groups. In some implementations, more fair includes improved accuracy by compensating for training data that under represents one or more constituent groups or elements of the training data. (See, Abstract; Para. 4-5, 91, 97; Fig. 1, 5));
A non-transitory computer readable storage medium storing instructions for mitigating disparities between outputs of different artificial intelligence / machine learning (AI/ML) models, the storage medium comprising executable code which, when executed by a processor, causes the processor to: (Ward, obtaining output from a machine learning (ML) model responsive to input data, obtaining initial training data representing training data used to train the ML model; fair includes having less disparate impact on protected versus unprotected groups. In some implementations, more fair includes improved accuracy by compensating for training data that under represents one or more constituent groups or elements of the training data. (See, Abstract; Para. 4-5, 91, 97; Claim 17; Fig. 1, 5));
receiving a first uncertainty value and a first set of model quality-related parameter values that are associated with a first model that is configured to generate a first loan price for a loan applicant; (Ward, a correction ML model configured to receive, as input, the input data and to output correction values which, when combined with the output from the ML model, perform the desired alteration, and generating corrected output as a combination of the output from the ML model and the output correction values from the correction ML model, and providing, for display, the corrected output. (See, Abstract); Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model; Adversarially debiased methods can include training an adversarial network to determine if a first model is making decisions that are correlated with specific features of a dataset (e.g., making loan decisions that have a disparate impact….Models generating the data shown in FIG. 4A can be configured to provide a probability of default classification for consumer loans. Initial data, e.g., the input features 102, can include a national sample of credit data with BISG labels. (See, Abstract; Para. 32-35, 78-80, 83-85, 91-93 Fig. 1, 4A));
receiving a second uncertainty value and a second set of model quality-related parameter values that are associated with a second model that is configured to generate a second loan price for the loan applicant; (Ward, the correction model 106 is a forest that includes one or more tree models. For example, one or more trees can be added for each iteration of training by the system 100; a correction model is trained to match error distribution instead of score distributions. As described in FIG. 1 , the system 100 can train the correction model 106 to generate corrections for the model 104 using differences in score values. In some implementations, the score correction engine 108 provides error distribution data to the correction model 106 as ground truth data (See, Abstract; Para. 24-25, 30, 32-35, 78-80, 83-85, 91-93 Fig. 1, 3, 4A)
receiving a set of feature weight functions that relate to weights of target metrics; (Ward, generate a model for each of gender and race. In the context of lending, both the ADV and correction models help to ensure that the given model output indicating loan decisions or terms have less of a disparate impact on one sex or another or one race or another…..the system 100 can train the correction model 106 using the error distribution data such that a combination of output from the correction model 106 and the model 104 for protected versus unprotected (or other subgroup pairings) results in an error distribution that is more equal between two or more groups than other models or algorithms. (See, Abstract; Para. 32-35, 78-80, 83-85, 91-93 Fig. 1, 4A)
calculating, for the first model based on the first uncertainty value, the first set of model quality-related parameter values, and the set of feature weight functions, a set of first model weights; (Ward, The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values. The model 104 and the correction model 106 can be any suitable type of machine learning model. Graphical representations of connected layers are shown for both the model 104 and the correction model 106 in FIG. 1; The process 200 includes generating a correction model using the correction training data (208). For example, the system 100 can generate the correction model 106. The system 100 can use the input features 102 as input to an untrained model (or partially trained) and compare the output of the untrained model to correction training data ground truths (e.g., data generated by the score correction engine 108 indicating difference values for one or more elements of the input features 102). Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
calculating, for the second model based on the second uncertainty value, the second set of model quality-related parameter values, and the set of feature weight functions, a set of second model weights; (Ward, the correction model 106 is a forest that includes one or more tree models. For example, one or more trees can be added for each iteration of training by the system 100….The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values. The model 104 and the correction model 106 can be any suitable type of machine learning model. Graphical representations of connected layers are shown for both the model 104 and the correction model 106 in FIG. 1; The process 200 includes generating a correction model using the correction training data (208). For example, the system 100 can generate the correction model 106. The system 100 can use the input features 102 as input to an untrained model (or partially trained) and compare the output of the untrained model to correction training data ground truths (e.g., data generated by the score correction engine 108 indicating difference values for one or more elements of the input features 102). Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 24, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
selecting a customized model based on the set of first model weights and the set of second model weights; (Ward, the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values; Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 24, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
receiving [a first tabular set of personal data] that comprises a first data subset that relates to individualized financial information associated with each respective applicant from among a plurality of loan applicants and a second data subset that relates to individualized demographic information associated with each respective applicant from among the plurality of loan applicants; (Ward, , scores generated by the model 104 are represented by graphs or charts, e.g., graph 302, 306 or chart 310. The model 104 provides generated scores to the score correction engine 108. The score correction engine 108 obtains protected class information 110. In some implementations, the system 100 obtains the protected class information 110 using one or more algorithms. For example, the system 100 can determine the protected class information 110, indicating a protected class corresponding to one or more persons or elements represented in the input feature 102; receiving personal data about a number of people, their credit scores from a black box model, and retrieving credit bureau data from a credit bureau, including tradeline data, inquiries, collections and public records. (See, Abstract; Para. 20-24, 56, 63; Fig. 1, 3));
training the customized model by using [the first tabular set of personal data], the target metrics, a predetermined set of sensitive labels, and historical information that relates to outputs generated by at least one from among the first model, the second model, and the customized model; (Ward, The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values….. a model that produces predictions that result in similar outcomes for a plurality of segments identified based on a sensitive attribute…. generate a prediction system that outputs one or more predictions based on received data (sometimes called predictive variables, model features, or independent variables), in a manner that produces similar outcomes with respect to different sensitive attributes. A method disclosed herein can enable model developers to correct a model for fairness by training a secondary correction model, such as correction model 106 shown in FIG. 1 , that modifies an original model's score or other output to make model-based outcomes more fair with respect to sensitive attributes….. the correction model 106 is a forest that includes one or more tree models. For example, one or more trees can be added for each iteration of training by the system 100; The correction model can be adjusted. In some implementations, the correction model is a forest model. For example, the correction model can be adjusted by adding a tree configured to provide a correction value for a particular set of input values. (See, Para.19-24; 34-36));
calculating an updated uncertainty value and an updated set of model quality-related parameter values for the trained customized model; and (Ward, The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values. The model 104 and the correction model 106 can be any suitable type of machine learning model. Graphical representations of connected layers are shown for both the model 104 and the correction model 106 in FIG. 1; The process 200 includes generating a correction model using the correction training data (208). For example, the system 100 can generate the correction model 106. The system 100 can use the input features 102 as input to an untrained model (or partially trained) and compare the output of the untrained model to correction training data ground truths (e.g., data generated by the score correction engine 108 indicating difference values for one or more elements of the input features 102). Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
using the trained customized model to generate a customized loan price for a first loan applicant from among the plurality of loan applicants. (Ward, Fairness of a prediction system can be identified based on a measured difference in outcomes for decisions based on the model for each value of a sensitive attribute. For example, if a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. Transmission of sensitive data can be secured using Transport Layer Security (TLS) or other transmission layer security. If the score exceeds a predetermined threshold, the loan application can be approved…. In some implementations, if the score otherwise satisfies a threshold, the loan application can be approved. In some implementations, if an application is neither auto-approved or auto-denied, it can be routed to underwriters for further processing. In this case, the score and reason codes can be displayed via a user interface so that an underwriter can consider them as they work with the applicant and determine whether they should be approved (See, Para. 19, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract)).
However, Ward does not specifically teach first tabular set of personal data.
Wu teaches the following limitation:
first tabular set of personal data; (Wu, the set of data features refer to a collection of properties or attributes that characterize the set of input data. In the case of tabular data, the set of data features may include a set of numerical features or a set of categorical features. As an example, in the case that the set of input data includes personal information provided by a loan applicant for use in determining their eligibility to qualify for a loan, the set of data features may include numerical features such as the age of the applicant, the income of the applicant, or the like, as well as categorical features such as the gender of the applicant, the educational level of the applicant, the occupation of the applicant, or the like. (See, Para. 64 and 93; Fig. 2 and 5)).
It would have been obvious to one of ordinary skill in the art before the effective filing of the claimed invention to have modified Ward with the features of Wu’s system because “loan applications are one area in which CFML techniques can be leveraged. For instance, in a loan application for an individual, a machine learning classifier may use information for an applicant (features such as income, educational background, age, marital status) to determine whether to approve or deny a loan. In the case that the applicant is denied a loan, the applicant may wish to know what they could do to increase their likelihood of being approved for a loan in the future. Here, counterfactual machine learning techniques may be used to analyze the relationships between the input features characterizing the applicant and the result, and provide recommendations to the applicant to increase their likelihood of being approved for a future loan application. As an example, a counterfactual machine learning model may provide a recommendation of “Your application is likely to be approved if you increase your income by 10%.” In this way, by using CFML techniques, users can gain valuable insights that assist them in a variety of decision-making scenarios.” (Wu, Para. 5).
Claims 2, 11, and 20:
Ward discloses:
wherein the predetermined set of sensitive labels includes at least one from among a race, a gender, and an age. (Ward, if a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. (See, Para.24, 35, 36; Fig. 1)).
Claims 3 and 12:
Ward teaches:
wherein each respective set of model quality-related parameters includes at least one from among a target loan price, a disparity, a robustness, and a stability. (Ward, in the context of lending, both the ADV and correction models help to ensure that the given model output indicating loan decisions or terms have less of a disparate impact on one sex or another or one race or another; (See, Para. 91); if a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. (See, Para. 34-36); models that can be used in fields with laws preventing disparate impact among groups, models that can be trained using training data that does not adequately represent a portion of constituents, among others. Methods described herein provide more fair models. More fair models can include models that are corrected to reduce disparate impact on one or more constituents in training or input data. In applications of lending, corrected models can have less of a disparate impact on marginalized communities compared to uncorrected credit lending models. (See, Para. 4, 20, 25, 83)).
Claims 4 and 13:
Ward teaches:
wherein the selecting of the customized model comprises linearly combining the first model with the second model. (Ward, The process 200 includes generating corrected output as a combination of the output from the model and output from the correction model (210). For example, the system 100, or another system, can use the correction model 106 with the model 104 to generate corrected output. The system 100 can use one or more combination techniques, e.g., linear combination, linear weighted combination, among others, to combine the output of the model 104 with output of the correction model 106. The combination of the output of the model 104 with output of the correction model 106 can be used as output of a corrected model. (See, Abstract; Para. 33, 51-54; Fig. 1)).
Claims 5 and 14:
Ward teaches:
wherein the individualized financial information includes at least one from among a credit score of each respective applicant, a loan-to-value ratio, and a loan principal amount. (Ward, receiving personal data about a number of people, their credit scores from a black box model, and retrieving credit bureau data from a credit bureau, including tradeline data, inquiries, collections and public records. (See, Abstract; Para. 35-36, 56, 73; Fig. 1)).
Claims 6 and 15:
Ward teaches:
wherein the individualized demographic information includes at least one from among a race of each respective applicant, a gender of each respective applicant, and an age of each respective applicant. (Ward, if a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. (See, Para.24, 35, 36; Fig. 1)).
Claims 7 and 16:
Ward teaches:
wherein the first model is configured to generate the first loan price based on a business-as-usual (BAU) paradigm that is designed to maximize profit and minimize financial loss without consideration of demographic fairness. (Ward, method can include recording approval rate, predictive accuracy, default risk, and business metrics such as gross profit and loss, disaggregated by protected status, for each model variation. (See, Para. 70)).
Claims 8 and 17:
Ward teaches:
wherein the second model is configured to generate the second loan price based on a demographic fairness paradigm that is designed to maximize profit and minimize financial loss while simultaneously ensuring that at least one metric that relates to demographic fairness is satisfied. (Ward, deploy a model (e.g., a predictive model) in a real-world scenario that impacts people's lives, fairness in how such a model impacts people's lives can be a consideration in determining whether to deploy the model, or continue model development. For example, whether a model favors a certain class of people (e.g., a class based on race, ethnicity, age, sex, national origin, sexual orientation, demographics, military status, etc.) over other classes of people may be a consideration in determining whether to deploy the model. (See, Para.14, 24); a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. (See, Para. 35-36)).
Claims 9 and 18:
Ward teaches:
further comprising: receiving a third uncertainty value and a third set of model quality-related parameter values that are associated with a third model that is configured to generate a third loan price for the loan applicant; and (Ward, Fairness of a prediction system can be identified based on a measured difference in outcomes for decisions based on the model for each value of a sensitive attribute. For example, if a prediction system is designed to predict the probability of default for a consumer loan, a fair model can generate output (approval rate for loans, for example) that is similar for each race and ethnicity (or other sensitive attribute such as age, color, disability, gender, gender expression, gender identity, genetic information, national origin, race, religion, sex, sexual orientation, marital status, or veteran status) of a credit applicant. Output can include an approve/deny decision and an approval rate. Output, also referred to as outcomes, can include loan terms (APR, down payment amount required) or any suitable or measurable outcome. Transmission of sensitive data can be secured using Transport Layer Security (TLS) or other transmission layer security. If the score exceeds a predetermined threshold, the loan application can be approved…. In some implementations, if the score otherwise satisfies a threshold, the loan application can be approved. In some implementations, if an application is neither auto-approved or auto-denied, it can be routed to underwriters for further processing. In this case, the score and reason codes can be displayed via a user interface so that an underwriter can consider them as they work with the applicant and determine whether they should be approved (See, Para. 19, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
calculating, for the third model based on the third uncertainty value, the third set of model quality-related parameter values, and the set of feature weight functions, a set of third model weights, (Ward, the correction model 106 is a forest that includes one or more tree models. For example, one or more trees can be added for each iteration of training by the system 100….The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values. The model 104 and the correction model 106 can be any suitable type of machine learning model. Graphical representations of connected layers are shown for both the model 104 and the correction model 106 in FIG. 1; The process 200 includes generating a correction model using the correction training data (208). For example, the system 100 can generate the correction model 106. The system 100 can use the input features 102 as input to an untrained model (or partially trained) and compare the output of the untrained model to correction training data ground truths (e.g., data generated by the score correction engine 108 indicating difference values for one or more elements of the input features 102). Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 24, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
wherein the selecting of the customized model is further based on the set of third model weights, and (Ward, the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values; Based on comparing the output of the untrained model to correction training data ground truths, the system 100 can adjust one or more weights or parameters of the untrained model to generate the correction model 106 as a trained version of the initial untrained model. (See, Para. 19, 24, 27, 31-36; 52, 78-80, 83-85; Fig. 1; Abstract));
wherein the training of the customized model is performed by using additional historical information that relates to outputs generated by the third model. (Ward, The model 104 or the correction model 106 can include one or more fully or partially connected layers of one or more nodes representing one or more weights or parameter values….. a model that produces predictions that result in similar outcomes for a plurality of segments identified based on a sensitive attribute…. generate a prediction system that outputs one or more predictions based on received data (sometimes called predictive variables, model features, or independent variables), in a manner that produces similar outcomes with respect to different sensitive attributes. A method disclosed herein can enable model developers to correct a model for fairness by training a secondary correction model, such as correction model 106 shown in FIG. 1 , that modifies an original model's score or other output to make model-based outcomes more fair with respect to sensitive attributes….. the correction model 106 is a forest that includes one or more tree models. For example, one or more trees can be added for each iteration of training by the system 100; The correction model can be adjusted. In some implementations, the correction model is a forest model. For example, the correction model can be adjusted by adding a tree configured to provide a correction value for a particular set of input values. (See, Para.19-24; 34-36)).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure is the following:
Brown (U.S. Patent Application Publication No. US 2023/0011777 A1) “System and method for property condition analysis”
Hill (U.S. Patent Application Publication No. US 2023/0087204 A1 ) “Systems and methods to screen a predictive model for risks of the predictive model”
Kamkar (U.S. Patent Application Publication No. US 2023/0105547 A1) “Machine learning model fairness and explainability”
Brugere (U.S. Patent Application Publication No. US 2024/0127331 A1) “Method and system for improving model fairness by using explainability techniques”
Bucklin (U.S. Patent Application Publication No. US 2024/0193481 A1 ) “Methods and systems for identification and visualization of bias and fairness for machine learning models”
Kamkar (U.S. Patent Application Publication No. US 2023/0377037 A1) “Systems and methods for generating gradient-boosted models with improved fairness”
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MOHAMMED H MUSTAFA whose telephone number is (571)270-7978. The examiner can normally be reached M-F 8:00 - 5:00.
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, Michael W Anderson can be reached on 571-270-0508. 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.
/MOHAMMED H MUSTAFA/Examiner, Art Unit 3693
/Mike Anderson/ Supervisory Patent Examiner, Art Unit 3693