Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 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 9-10 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 applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 9 recites the limitation "a second generative adversarial network" in Line 3. There is insufficient antecedent basis for this limitation in the claim. Claim 9 depends from claim 8, which depends from claim 4, and neither claims 4 and 8, or any claim from which claim 4 depends from, recites a corresponding “first generative adversarial network”. It is unclear if the recited “second generative adversarial network” requires the existence of a first generative adversarial network or if the recited “second generative adversarial network” is intended to distinguish from a “first generative adversarial network”. Thus, claim 9 is indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor regards as the invention. Claim 10 is rejected as being indefinite for similar reasons.
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 is/are rejected under 35 U.S.C. 101 because they are directed to an abstract idea without significantly more.
Regarding claims 1-20,
Step 1: The preamble of claims 1-14 recites a method, which falls within the statutory category of a process. The preamble of claims 15-19 recites a device, which falls within the statutory category of an apparatus. The preamble of claim 20 recites a non-transitory computer readable storage medium, which falls within the statutory category of a manufacture.
Regarding claim 1,
Step 2A – Prong One: Claim 1 recites:
A method for determining fairness impact of a sample on a model, comprising:
generating a counterfactual sample by adjusting an original sample in an original sample set, the original sample set being used for generating an original model for performing a classification task;
determining a fairness metric of the original model on a validation sample set; and
determining fairness impact of the original sample on the original model based on the fairness metric, the original sample, and the counterfactual sample.
The broadest reasonable interpretation, in light of the Specification, of the bolded limitations above are directed to a mental process able to be performed in the human mind or by using pen and paper. A human could use observation, evaluation, and judgement to generate a counterfactual sample by adjusting an original sample in an original sample set, determine a fairness metric of the original sample, and determine a fairness impact of the original sample based on the determined fairness metric, original sample, and counterfactual sample.
Step 2A – Prong One (Yes).
Step 2A – Prong Two: There are no additional elements in the claim that would integrate the judicial exception into a practical application.
Step 2A – Prong Two (No).
Step 2B: There are no additional elements in the claim that would amount to significantly more than the judicial exception.
Step 2B (No).
Claim 1 is ineligible.
Regarding claims 15 and 20,
These claims are similar in scope to claim 1 and are rejected under similar rationale. The processors and memory in these claims are generic computing components (See MPEP 2106.05(f)).
Claims 15 and 20 are ineligible.
Dependent claims:
Claims 2, 4-5, 8, 11, 16, and 18-19: These claims recite further abstract ideas (mental processes) and thus are ineligible.
Claims 3 and 17: These claims recite further abstract ideas (mental processes and mathematical concepts). Generating a plurality of losses corresponding to a plurality of original samples is a mental process that can be performed in the human mind or by using pen and paper. Determining a Hessian matrix based on the plurality of losses is a mathematical relationship and thus falls within the mathematical concepts grouping of abstract ideas. Determining the fairness impact value using the Hessian matrix, the fairness metric, the first loss, and the second loss is a mental process that can be performed in the human mind or by using pen and paper. A human could use observation, evaluation, and judgement to generate a plurality of losses and determine a fairness impact value using the Hessian matrix, the fairness metric, the first loss, and the second loss. There are no additional elements in the claim that would integrate the judicial exception into a practical application, and there are no additional elements in the claim that would amount to significantly more than the judicial exception. Thus, these claims are ineligible.
Claims 6 and 9: These claims recite additional elements regarding “generating… a second feature set… by using a first generative adversarial network” and “generating… a label of the counterfactual sample by using the original model” in claim 6, and “generating… a second feature set… by using a second generative adversarial network” and “generating… a label of the counterfactual sample by using the original model” in claim 9. These additional elements are mere instructions to apply the judicial exception on a generic computer (See MPEP 2106.05(f)). The computer is recited at a high level of generality and imposes no meaningful limitations on the claim. These additional elements fail to integrate the judicial exception into a practical application, and these additional elements fail to amount to significantly more than the judicial exception. These claims are ineligible.
Claims 7 and 10: These claims recite further abstract ideas (mental processes, “determining a first original sample subset…” and “generating a target sample based on…”). The additional element of these claims regarding “obtaining a first population sample…” in claim 7 and “obtaining a first feature sample…” in claim 10, are insignificant extra-solution activities that amount to mere data gathering (See MPEP 2106.05(g)). Data gathering is a well-understood, routine conventional activity as recognized by the courts (See MPEP 2106.05(d)(II)). The additional elements of these claims regarding “generating the first generative adversarial network based on…” in claim 7 and “generating the second generative adversarial network based on…” in claim 10 are mere instructions to apply the judicial exception on a generic computer (See MPEP 2106.05(f)). The computer is recited at a high level of generality and imposes no meaningful limitations on the claim. These additional elements fail to integrate the abstract idea into a practical application, and these additional elements fail to amount to significantly more than the judicial exception. These claims are ineligible.
Claims 12-14: These claims recite further abstract ideas (mental processes, “determining a plurality of fairness impact values…”, “determining… a plurality of target samples…”, “generating an updated sample set by changing the labels…”, and “generating an updated sample set by removing the plurality of target samples…”). The additional element of claim 14 regarding “resampling an updated sample in response to a proportion of fairness impact values greater than…” is insignificant extra-solution activity that amounts to mere data gathering (See MPEP 2106.05(g)). Data gathering is well-understood, routine conventional activity as recognized by the courts (See MPEP 2106.05(d)(II)). The additional elements of these claims regarding “generating an updated model based on the updated sample set” are mere instructions to apply the judicial exception on a generic computer (See MPEP 2106.05(f)). The computer is recited at a high level of generality and imposes no meaningful limitations on the claim. Thus, the additional elements fail to integrate the judicial exception into a practical application and the additional elements fail to amount to significantly more than the judicial exception. These claims are ineligible.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-6 and 14- 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione et al. (US Pub. No. 2022/0114399, published April 2022, hereinafter “Castiglione”) in view of Li et al. (NPL: Achieving Fairness at No Utility Cost via Data Reweighting with Influence, published July 2022, hereinafter “Li”).
Regarding claim 1, Castiglione teaches a method for determining fairness impact of a sample on a model, comprising:
generating a counterfactual sample by adjusting an original sample in an original sample set, the original sample set being used for generating an original model for performing a classification task (Castiglione, [0083] – “X: Feature (input) variables. When features are observed, Applicants use x to represent the feature vector.”, [0084] – “Y: Prediction (output) variables. When a label is observed, Applicants use y to represent the label as a scalar. As Applicant can conduct fairness testing on binary classification tasks, for example the prediction y can be a probability.” and in [0175] – “Counterfactual Fairness (CFF), like FTA, mandates that similar individuals be treated similarly, but it does so causally. More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c.” – teaches generating a counterfactual sample (in Castiglione, the counterfactual sample is x’) by adjusting an original sample (x) in an original sample set (X), the original sample set being used for generating an original model for performing a classification task (conducts fairness testing on binary classification tasks, outputs y can be a classification probability));
determining a fairness metric of the original model (Castiglione, [0149] – “Systems and methods are provided which may allow testing and diagnosing a machine learning model for fairness and which may generate a fairness indicator value representing how fair the machine learning model is… The system may generate output data representative of the fairness indicator value. The fairness indicator value may be indicative of discrimination risk due to the target predictions generated by the machine learning model.” – teaches determining a fairness metric (fairness indicator value) of the original model (fairness indicator value indicates how fair the model is)); and
determining fairness impact of the original sample on the original model based on the fairness metric, the original sample, and the counterfactual sample (Castiglione, [0261, 0262] – teaches determining a fairness impact (generalized local fairness) of the original sample on the original model based on the fairness metric, the original sample, and the counterfactual sample (generalized local fairness of original sample x on the original model f determined using the fairness indicator value δ, the original sample x, and the counterfactual sample x’)).
Castiglione fails to explicitly teach determining a fairness metric of the original model on a validation sample set.
However, analogous to the field of the claimed invention, Li teaches:
determining a fairness metric of the original model on a validation sample set (Li, Section 3 Paragraph 6 – “Instead of evaluating the fairness and utility on the training set T , we are interested in the change of this two-side performance for the classifier on a validation set V = {zj = (xj,yj)}Nv j=1 before and after a reweighing over T , where each sample zj in the validation set is associated with a binary sensitive attribute aj ∈ {0,1}.” and in Section 3 Paragraph 8 – “Here we use a truncated subscript ‘eop/dp’ to express the feasibility for either functions, and in what follows we may use fVfair to unify fVeop/dp.” – teaches determining a fairness metric of the original model on a validation sample set (determines fairness metric ffairV where V is a validation set));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the determination of a fairness metric of the model on a validation sample set of Li to the counterfactual samples, models, and fairness impact determination of Castiglione. Doing so would directly diagnose and correct sources of bias (Li, Introduction) and help identify which sample has a positive impact on training, and to which direction and what extent (Li, Section 3 Paragraph 6).
Claims 15 and 20 incorporate substantively all the limitations of claim 1 in a device and non-transitory computer readable storage medium, and are rejected on similar grounds as above. Castiglione teaches the processors and memory of these claims at [0074] – “The approaches described herein are practically implemented in the form of computing devices or servers that include computer processors, memory, and data storage, and methods that operate thereon or are residing as machine readable instruction sets stored in non-transitory computer readable media.”
Regarding claim 2, the combination of Castiglione and Li teaches the method according to claim 1,
determining, based on the original sample, a first loss using a loss function of the original model (Li, Section 3 Paragraph 4 – “To formulate the fairness issue as an optimization problem, one can intuitively quantify the inequality and turn the subjection of fairness into an objective function. For instance, with ℓ0/1 denotes zero-one loss, the gap in Equal Opportunity is: Eq. (7)” and in Section 3 Paragraph 5 – “To make the function differentiable, a surrogate function is necessary to replace ℓ0/1 and here we substitute it with the training loss ℓ for Equal Opportunity, following previous works (Zafar et al., 2017; Donini et al., 2018). Demographic Parity can be derived similarly. The fair loss over a sample set S for these two notions are: Eq. (8)” – teaches determining, based on the original sample, a first loss using a loss function of the original model (in Eq. (8) the first term of feopS is a loss using a loss function of the original model based on the original sample, where a=1));
determining, based on the counterfactual sample, a second loss using the loss function (Li, Section 3 Paragraph 4 – “To formulate the fairness issue as an optimization problem, one can intuitively quantify the inequality and turn the subjection of fairness into an objective function. For instance, with ℓ0/1 denotes zero-one loss, the gap in Equal Opportunity is: Eq. (7)” and in Section 3 Paragraph 5 – “To make the function differentiable, a surrogate function is necessary to replace ℓ0/1 and here we substitute it with the training loss ℓ for Equal Opportunity, following previous works (Zafar et al., 2017; Donini et al., 2018). Demographic Parity can be derived similarly. The fair loss over a sample set S for these two notions are: Eq. (8)” – teaches determining, based on the original sample, a first loss using a loss function of the original model (in Eq. (8) the second term of feopS is a loss using a loss function of the original model based on the original sample, where a=0)); and
determining a fairness impact value indicative of the fairness impact based on the fairness metric, the first loss, and the second loss (Li, Section 3 Paragraph 6 – “Instead of evaluating the fairness and utility on the training set T , we are interested in the change of this two-side performance for the classifier on a validation set V = {zj = (xj,yj)}Nv j=1 before and after a reweighing over T , where each sample zj in the validation set is associated with a binary sensitive attribute aj ∈ {0,1}.”, Section 3 Paragraph 7 – “The influence function on fairness can be derived by realizing the function f in Equation (4) with Equation (8): Eq. (9)” and in Section 3 Paragraph 8 – “Here we use a truncated subscript ‘eop/dp’ to express the feasibility for either functions, and in what follows we may use fVfair to unify fVeop/dp.” – teaches determining a fairness impact value indicative of the fairness impact based on the fairness metric, the first loss, and the second loss (determines an influence function on fairness based on the first and second losses that appear in feopS in Eq. (8) and the fairness metric ffairV, where ffairV is used to unify feop/dpV, where V is a validation set)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the fairness impact value, first loss, and second loss determinations of Li to further modify the fairness impact determination of Castiglione and Li. Doing so would help identify which sample has a positive impact on training, and to which direction and what extent (Li, Section 3 Paragraph 6).
Claim 16 is similar to claim 2, hence similarly rejected.
Regarding claim 3, the combination of Castiglione and Li teaches the method according to claim 2, wherein determining the fairness impact value comprises:
generating a plurality of losses corresponding to a plurality of original samples in the original sample set by using the loss function (Li, Section 2 Paragraph 4 – “Influence function measures the effect of changing an infinitesimal weight from samples in w, then linearly extrapolates to complete all of w. It assumes ℓ to be twice differentiable and strictly convex in θ, and f to be differentiable as well. These assumptions are mild and feasible for many pipelines involving classifiers like logistic regression to process tabular data. Having θ(1) and θ(1 − w) satisfied their first-order optimality conditions, and by taking a Taylor approximation, the actual influence I∗ f(w) can be approximated by an estimation Eq. (4) where Hˆθ(1) = ∑i=1NT ∇θ2ℓ(zi; ˆθ(1)) is the Hessian matrix of ℓ, and the convexity ensures its invertibility.” – teaches generating a plurality of losses corresponding to a plurality of original samples in the original sample set by using the loss function (takes summation from 1 to NT, where NT is the size N of training set T, of loss l on samples zi, thus generating a plurality of losses corresponding to a plurality of original samples in the original sample set by using the loss function));
determining a Hessian matrix for the original model based on the plurality of losses (Li, Section 2 Paragraph 4 – teaches determining a Hessian matrix for the original model based on the plurality of losses (in Eq. (4), H^ θ(1) is the Hessian matrix of the plurality of losses l, thus determining a Hessian matrix for the original model based on the plurality of losses)); and
determining the fairness impact value based on the Hessian matrix, the fairness metric, the first loss, and the second loss (Li, Section 3 Paragraph 7 – “The influence function on fairness can be derived by realizing the function f in Equation (4) with Equation (8): Eq. (9)” and in Section 3 Paragraph 8 – “Here we use a truncated subscript ‘eop/dp’ to express the feasibility for either functions, and in what follows we may use fVfair to unify fVeop/dp.” – teaches determining the fairness impact value based on the Hessian matrix (in Eq. (9), the Hessian matrix is denoted as H^θ(1)-1), the fairness metric (ffairV, where ffairV is used to unify feop/dpV), the first loss, and the second loss (first and second loss of feopS, which appear in Eq. (8))).
Claim 17 is similar to claim 3, hence similarly rejected.
Regarding claim 4, the combination of Castiglione and Li teaches the method according to claim 1, wherein adjusting the original sample comprises one of:
changing a population attribute of the original sample, changing a feature of the original sample, changing a label of the original sample, or removing the original sample from the original sample set (Castiglione, [0036] – “In an alternate embodiment, the specialized computing system is utilized as an auditor or adjudicator subsystem that receives input models, and is configured to automatically flag or cause the models to be re-trained, re-engineered, or removed from service for an estimated fairness violation”, [0037] – “Re-engineering can include changing the structures of the models or removing features or nodes from analysis from the models (e.g., changing the possible input feature sets or removing nodes from the latent space).”, [0098] – “Given observed feature x, prediction y and protected variables c, counterfactual fairness holds if: P(y|x,c)=P.sub.C←do(c′)(y|x,c)”, and in [0099] – “where the do(⋅) operation (Pearl et al. 2009) changes protected attributes from c to c′ during inference.” – teaches wherein adjusting the original sample comprises one of changing a population attribute of the original sample (attribute c becomes c’, thus changing a population attribute of the original sample) or removing the original sample from the original sample set (automatically flags or causes models to be re-engineered, re-engineering includes removing features or nodes from latent space)).
Claim 18 is similar to claim 4, hence similarly rejected.
Regarding claim 5, the combination of Castiglione and Li teaches the method according to claim 4, wherein generating the counterfactual sample comprises:
in response to the adjustment for the original sample being changing the population attribute of the original sample, generating the counterfactual sample based on the changed population attribute and a first feature set of the original sample (Castiglione, [0098] – “Given observed feature x, prediction y and protected variables c, counterfactual fairness holds if: P(y|x,c)=P.sub.C←do(c′)(y|x,c)”, and in [0099] – “where the do(⋅) operation (Pearl et al. 2009) changes protected attributes from c to c′ during inference.” – teaches wherein generating the counterfactual sample comprises: in response to the adjustment for the original sample being changing the population attribute of the original sample, generating the counterfactual sample based on the changed population attribute and a first feature set of the original sample (changes protected attributes from c to c’ during inference, thus generating a counterfactual sample based on the changed population attribute and first feature set of the original sample)).
Claim 19 is similar to claim 5, hence similarly rejected.
Regarding claim 6, the combination of Castiglione and Li teaches the method according to claim 5, wherein generating the counterfactual sample further comprises:
generating, based on the changed population attribute and the first feature set of the original sample, a second feature set of the counterfactual sample by using a first generative adversarial network (Castiglione, [0164] – “Given a datapoint (x,y), an adversarial example x′ is obtained as a solution of a constrained optimization problem: Eq. (3)”, [0165] – “In this way, x′ is a point in the neighbourhood of x for which outputs the model ƒ.sub.Y change rapidly. Both exhaustive and approximate algorithms for finding adversarial examples may be obtained.”, [0175] – “More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c. ” and in [0176] – “In particular, equation (7) may fit within the more general framework of model-based robust deep learning, where robustness is enforced in latent spaces as opposed to the raw inputs. In practice, however, CFF is hard to implement, because it requires access to a causal graph. FlipTest attempts to relax this requirement by generating counterfactuals using generative adversarial networks.” – teaches generating, based on the changed population attribute (c’) and the first feature set of the original sample (feature x of original sample), a second feature set (x’) of the counterfactual sample by using a first generative adversarial network (utilizes Generative Adversarial Networks to generate counterfactual features x’)); and
generating, based on the second feature set of the counterfactual sample, a label of the counterfactual sample by using the original model (Castiglione, [0084] – “Y: Prediction (output) variables. When a label is observed, Applicants use y to represent the label as a scalar. As Applicant can conduct fairness testing on binary classification tasks, for example the prediction y can be a probability. ” and in [0254] – “FIG. 5A and FIG. 5B show an example of generating fair counterfactuals using constrained optimization. When a model is approximately constant near a point x, a single perturbation using the gradient may fail to reveal nearby discriminatory behaviour. One can thus use an iterative algorithm to find points x′ which maximize the change in the model's outputs, subject to constraints.” – teaches generating, based on the second feature set of the counterfactual sample (x’), a label (y) of the counterfactual samples by using the original model (finds points x’ which maximize change in model outputs, model outputs are labels represented as scalars, thus teaching generating a label of the counterfactual sample by using the original model)).
Regarding claim 14, the combination of Castiglione and Li teaches the method according to claim 1, further comprising:
determining a plurality of fairness impact values corresponding to the plurality of original samples by changing population attributes of the plurality of original samples in the original sample set respectively (Castiglione, [0175] – “Counterfactual Fairness (CFF), like FTA, mandates that similar individuals be treated similarly, but it does so causally. More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c.”, [0261, 0262], and in [0329] – “In various embodiments, the (input) value is one of a plurality of values configured to be received by the one or more processors, the fairness indicator value is one of a plurality of fairness indicator values, each of the plurality of fairness indicator values generated based on a corresponding one of the plurality of values, and the output data is indicative of whether an aggregated measure of the plurality of fairness indicators exceeds a predefined fairness threshold.” – teaches determining a plurality of fairness impact values (plurality of generalized local fairness values generated using plurality of fairness indicator values) corresponding to the plurality of original samples by changing the population attributes of the plurality of original samples in the original sample set respectively (generalized local fairness values correspond to plurality of original samples x by changing the population attributes c of the plurality of original samples in the original sample set));
resampling an updated sample set from a different population in response to a proportion of fairness impact values greater than a predetermined value among the plurality of fairness impact values being greater than a predetermined proportion (Castiglione, [0329] – “In various embodiments, the (input) value is one of a plurality of values configured to be received by the one or more processors, the fairness indicator value is one of a plurality of fairness indicator values, each of the plurality of fairness indicator values generated based on a corresponding one of the plurality of values, and the output data is indicative of whether an aggregated measure of the plurality of fairness indicators exceeds a predefined fairness threshold.” and in [0517] – “In a variant embodiment, the system automatically takes action instead without human intervention (e.g., picking a fairest model for use, deactivating unfair models, having models retrained or perturbed automatically).” – teaches resampling an updated sample set from a different population in response to a proportion of fairness impact values greater than a predetermined value among the plurality of fairness impact value being greater than a predetermined portion (system may automatically retrain models in response to predefined fairness threshold being exceeded, thus resampling an updated sample set from a different population c in response to an proportion of fairness impact values greater than a predetermined value among the plurality of fairness impact values)); and
generating an updated model based on the updated sample set (Castiglione, [0036] – “ In an alternate embodiment, the specialized computing system is utilized as an auditor or adjudicator subsystem that receives input models, and is configured to automatically flag or cause the models to be re-trained, re-engineered, or removed from service for an estimated fairness violation.” and in [0037] – “Re-training can include resetting or randomizing model weights and re-conducting training, training with different training sets, among others (e.g., re-training a model using training set B instead of training set A). Re-engineering can include changing the structures of the models or removing features or nodes from analysis from the models (e.g., changing the possible input feature sets or removing nodes from the latent space).” – teaches generating an updated model based on the updated sample set (computing system utilized as an auditor that automatically flags models to be re-trained or re-engineered. Re-training includes training with different sets, re-engineering includes changing model structure or removing features/nodes from analysis from the models)).
Claim(s) 7 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione and Li as applied to claims 1, 15, and 20 above, and further in view of Black et al. (NPL: FlipTest: Fairness Testing via Optimal Transport, published Jan. 2020, hereinafter “Black:).
Regarding claim 7, the combination of Castiglione and Li teaches the method according to claim 6.
The combination of Castiglione and Li fails to explicitly teach further comprising: determining a first original sample subset and a second original sample subset from the original sample set, a population attribute of a sample in the first original sample subset having a first value, and a population attribute of a sample in the second original sample subset having a second value; obtaining a first population sample in the first original sample subset; generating a target sample based on a feature set of the first population sample and the second value of the population attribute; and generating the first generative adversarial network based on the target sample and the second original sample subset.
However, analogous to the field of the claimed invention, Black teaches: further comprising:
determining a first original sample subset and a second original sample subset from the original sample set, a population attribute of a sample in the first original sample subset having a first value, and a population attribute of a sample in the second original sample subset having a second value (Black, Section 3 Paragraph 2 – “We first introduce the notation. Let S and S′ be two distributions defined over the feature space X. In practice, we do not know these distributions, so we usually deal with observations of points drawn from these distributions instead. We will use the sets S = {x1,...,xn} and S′ = {x′ 1,...,x′ n} to denote the ob served points, where n = |S| = |S′|.” and in Section 4.1 Paragraph 2 – “In our experiments, S and S′ will correspond to two groups with differing values for a protected attribute, and h will be a classifier with the potential to be unfair.” – teaches determining a first original sample subset (S) and a second original sample subset (S’) from the original sample set (X), a population attribute of a sample in the first original sample subset having a first value, and a population attribute of a sample in the second original sample subset having a second value (S and S’ correspond to two groups with differing values for a protected attribute, thus determining a population attribute of a sample in the first original sample subset having a first value and a population attribute of a sample in the second original sample subset having a second value));
obtaining a first population sample in the first original sample subset (Black, Section 3.2 Paragraph 2 – “To measure stability, we fix a point x ∈ S and then draw multiple distinct samples of the other n − 1 points from S and n points from S′.” – teaches obtaining a first population sample in the first original sample subset (fixes a point x in the first original sample subset S));
generating a target sample based on a feature set of the first population sample and the second value of the population attribute (Black, Section 3.1 Paragraph 2 – “For the purpose of finding an optimal transport mapping, we modify the generator’s loss function to take into account the cost of moving from a point in S to a point in S′: Eq. (1)”, Section 3.2 Paragraph 2 – “To measure stability, we fix a point x ∈ S and then draw multiple distinct samples of the other n − 1 points from S and n points from S′. Thus, we have different sampled sets S and S′ each time, and we observe the variance of the point f (x) over the random draws.”, and in Section 4.1 Paragraph 2 – “In our experiments, S and S′ will correspond to two groups with differing values for a protected attribute, and h will be a classifier with the potential to be unfair.” – teaches generating a target sample based on a feature set of the first population sample (x) and the second value of the population attribute (S’ has a differing attribute value from S. Observes variance of point f(x) over random draws from S and S’, thus generating a target sample based on x and mapping from S to S’, where S’ has a different attribute value than S)); and
generating the first generative adversarial network based on the target sample and the second original sample subset (Black, Section 3.1 Paragraph 2 – “For the purpose of finding an optimal transport mapping, we modify the generator’s loss function to take into account the cost of moving from a point in S to a point in S′: Eq. (1)” and in Section 3.1 Paragraph 3 – ‘Our modified generator has two objectives, with the parameter λ controlling their relative importance: generating the correct output distribution S′, and minimizing the expected costc(x,G(x)). Proposition 1 formalizes the intuition that these objectives are also those of an optimal transport mapping.” – teaches generating the first adversarial network based on the target sample and the second original sample subset (generates GAN with optimal transport mapping utilizing target sample and second sample subset S’)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the sample subsets, population attribute values, and generation of a generative adversarial network of Black to the original samples, generative models, and fairness determinations of Castiglione and Li. Doing so would provide an optimal transport map to construct instances that reveal whether a model’s behavior is sensitive to changes in protected status and provide methods for approximating optimal transport based on generative adversarial networks (Black, Introduction).
Claim(s) 8-9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione and Li as applied to claims 1, 15, and 20 above, and further in view of Dalli et al. (US Patent No. 11,256,989, published Feb. 2022, hereinafter “Dalli”).
Regarding claim 8, the combination of Castiglione and Li teaches the method according to claim 4.
The combination of Castiglione and Li fails to explicitly teach wherein generating the counterfactual sample comprises: in response to the adjustment for the original sample being changing a feature of the original sample, generating the counterfactual sample based on the changed feature and the first feature set of the original sample.
However, analogous to the field of the claimed invention, Dalli teaches: wherein generating the counterfactual sample comprises:
in response to the adjustment for the original sample being changing a feature of the original sample, generating the counterfactual sample based on the changed feature and the first feature set of the original sample (Dalli, Pg. 15, Col. 6, Line 55 – Pg. 16, Col. 7, Line 15 – “Bias detection determines if a specific model or dataset is biased towards specific features. However, such features may depend on other features, which may result in a causal effect… Causal models may also be used in certain cases to correct or adjust for bias by creating what-if and what-if-not type of analyses and may also suggest suitable interventions. Causal models may also be used to detect the presence of bias using counterfactuals and may also determine how results from any of: (i.) one part of the model, (ii.) a related model, and/or (iii.) sub-part of the model, can be transferred or transposed directly or via the appropriate modification, conditioning, allowance or other appropriate transformation.” and in Pg. 16, Col. 7, Line 65 – Pg. 17, Col. 8, Line 3 – “Bias detection may be detected by analysing the difference in feature attribution when applying a control swap. Control swap is a method whereby for a given transaction X, the potential data field which has a bias (for example gender), is modified to test any change in the outcome and explanation.” – teaches wherein generating the counterfactual sample comprises: in response to the adjustment for the original sample being changing a feature of the original sample, generating the counterfactual sample based on the changed feature and the first feature set of the original sample (applies control swap to apply a difference in feature attribution of a sample, modifies data field of sample X to generate a counterfactual sample based on a changed feature and the feature of the original sample)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the feature modification to generate counterfactual samples of Dalli to the counterfactual samples, fairness metrics, and fairness impact determinations of Castiglione and Li. Doing so would provide methods for bias detection by analyzing differences in samples with modified features (Dalli, Pg. 16, Cols. 7-8).
Regarding claim 9, the combination of Castiglione, Li, and Dalli teaches the method according to claim 8, wherein generating the counterfactual sample comprises:
generating, based on the changed feature and the first feature set of the original sample, a second feature set of the counterfactual sample by using a second generative adversarial network (Castiglione, [0164] – “Given a datapoint (x,y), an adversarial example x′ is obtained as a solution of a constrained optimization problem: Eq. (3)”, [0165] – “In this way, x′ is a point in the neighbourhood of x for which outputs the model ƒ.sub.Y change rapidly. Both exhaustive and approximate algorithms for finding adversarial examples may be obtained.”, [0175] – “More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c. ” and in [0176] – “In particular, equation (7) may fit within the more general framework of model-based robust deep learning, where robustness is enforced in latent spaces as opposed to the raw inputs. In practice, however, CFF is hard to implement, because it requires access to a causal graph. FlipTest attempts to relax this requirement by generating counterfactuals using generative adversarial networks.” – teaches generating, based on the changed feature (changed feature of Dalli) and the first feature set of the original sample (feature x of original sample), a second feature set (x’) of the counterfactual sample by using a second generative adversarial network (utilizes Generative Adversarial Networks to generate counterfactual features x’)); and
generating, based on the second feature set of the counterfactual sample, a label of the counterfactual sample by using the original model (Castiglione, [0084] – “Y: Prediction (output) variables. When a label is observed, Applicants use y to represent the label as a scalar. As Applicant can conduct fairness testing on binary classification tasks, for example the prediction y can be a probability. ” and in [0254] – “FIG. 5A and FIG. 5B show an example of generating fair counterfactuals using constrained optimization. When a model is approximately constant near a point x, a single perturbation using the gradient may fail to reveal nearby discriminatory behaviour. One can thus use an iterative algorithm to find points x′ which maximize the change in the model's outputs, subject to constraints.” – teaches generating, based on the second feature set of the counterfactual sample (x’), a label (y) of the counterfactual samples by using the original model (finds points x’ which maximize change in model outputs, model outputs are labels represented as scalars, thus teaching generating a label of the counterfactual sample by using the original model)).
Claim(s) 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione, Li, and Dalli as applied to claims 1, 15, and 20 above, and further in view of Black.
Regarding claim 10, the combination of Castiglione, Li, and Dalli teaches the method according to claim 9.
The combination of Castiglione, Li, and Dalli fails to explicitly teach further comprising: determining a first original sample subset and a second original sample subset from the original sample set, the feature of a sample in the first original sample subset having a first value and the feature of a sample in the second original sample subset having a second value; obtaining a first feature sample in the first original sample subset; generating a target sample based on a feature set of the first feature sample and the second value of the feature; and generating the second generative adversarial network based on the target sample and the second original sample subset.
However, analogous to the field of the claimed invention, Black teaches:
determining a first original sample subset and a second original sample subset from the original sample set, the feature of a sample in the first original sample subset having a first value and the feature of a sample in the second original sample subset having a second value (Black, Section 3 Paragraph 2 – “We first introduce the notation. Let S and S′ be two distributions defined over the feature space X. In practice, we do not know these distributions, so we usually deal with observations of points drawn from these distributions instead. We will use the sets S = {x1,...,xn} and S′ = {x′ 1,...,x′ n} to denote the observed points, where n = |S| = |S′|.” and in Section 4.1 Paragraph 2 – “In our experiments, S and S′ will correspond to two groups with differing values for a protected attribute, and h will be a classifier with the potential to be unfair.” – teaches determining a first original sample subset (S) and a second original sample subset (S’) from the original sample set (X), the feature of a sample in the first original sample subset having a first value, and the feature of a sample in the second original sample subset having a second value (S and S’ correspond to two groups with differing feature values, thus determining a feature of a sample in the first original sample subset having a first value and a feature of a sample in the second original sample subset having a second value));
obtaining a first feature sample in the first original sample subset (Black, Section 3.2 Paragraph 2 – “To measure stability, we fix a point x ∈ S and then draw multiple distinct samples of the other n − 1 points from S and n points from S′.” – teaches obtaining a first feature sample in the first original sample subset (fixes a point x in the first original sample subset S));
generating a target sample based on a feature set of the first feature sample and the second value of the feature (Black, Section 3.1 Paragraph 2 – “For the purpose of finding an optimal transport mapping, we modify the generator’s loss function to take into account the cost of moving from a point in S to a point in S′: Eq. (1)”, Section 3.2 Paragraph 2 – “To measure stability, we fix a point x ∈ S and then draw multiple distinct samples of the other n − 1 points from S and n points from S′. Thus, we have different sampled sets S and S′ each time, and we observe the variance of the point f (x) over the random draws.”, and in Section 4.1 Paragraph 2 – “In our experiments, S and S′ will correspond to two groups with differing values for a protected attribute, and h will be a classifier with the potential to be unfair.” – teaches generating a target sample based on a feature set of the first feature sample (x) and the second value of the feature (S’ has a differing value from S. Observes variance of point f(x) over random draws from S and S’, thus generating a target sample based on x and mapping from S to S’, where S’ has a different attribute value than S)); and
generating the second generative adversarial network based on the target sample and the second original sample subset (Black, Section 3.1 Paragraph 2 – “For the purpose of finding an optimal transport mapping, we modify the generator’s loss function to take into account the cost of moving from a point in S to a point in S′: Eq. (1)” and in Section 3.1 Paragraph 3 – ‘Our modified generator has two objectives, with the parameter λ controlling their relative importance: generating the correct output distribution S′, and minimizing the expected costc(x,G(x)). Proposition 1 formalizes the intuition that these objectives are also those of an optimal transport mapping.” – teaches generating the first adversarial network based on the target sample and the second original sample subset (generates GAN with optimal transport mapping utilizing target sample and second sample subset S’)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the sample subsets, target samples, and generation of generative adversarial networks of Black to the feature modifications, original samples, and fairness determinations of Castiglione, Li, and Dalli. Doing so would provide an optimal transport map to construct instances that reveal whether a model’s behavior is sensitive to changes in protected status and provide methods for approximating optimal transport based on generative adversarial networks (Black, Introduction).
Claim(s) 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione and Li as applied to claims 1, 15, and 20 above, and further in view of Wang et al. (NPL: Robustness to Spurious Correlations in Text Classification via Automatically Generated Counterfactuals, published 2021, hereinafter “Wang”).
Regarding claim 11, the combination of Castiglione and Li teaches the method according to claim 4.
The combination of Castiglione and Li fails to explicitly teach: wherein generating the counterfactual sample comprises: in response to the adjustment for the original sample being changing a label of the original sample, generating a counterfactual sample corresponding to the original sample based on a feature set of the original sample and the changed label.
However, analogous to the field of the claimed invention, Wang teaches: wherein generating the counterfactual sample comprises:
in response to the adjustment for the original sample being changing a label of the original sample, generating a counterfactual sample corresponding to the original sample based on a feature set of the original sample and the changed label (Wang, Section “Problem and Motivation” Paragraph 4 – “Specifically, for a sample (d,y), we get the corresponding counterfactual sample (d,y) by (i) substituting causal terms in d with their antonyms to get d’ and(ii) assigning an opposite label y’ to d’.” – teaches wherein generating the counterfactual sample comprises: in response to the adjustment for the original sample being changing a label of the original sample, generating a counterfactual sample corresponding to the original sample based on a feature set of the original sample (d,y) and the changed label (opposite label y’)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the changing of labels of original samples of Wang to the original samples, models, and counterfactual sample determinations of Castiglione and Li. Doing so would provide methods for automatic generation of counterfactual samples with opposite labels to train robust classifiers (Wang, Introduction).
Claim(s) 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione and Li as applied to claims 1, 15, and 20 above, and further in view of Koh et al. (NPL: Understanding Black-box Predictions via Influence Functions, published Dec. 2020, hereinafter “Koh”) and Wang.
Regarding claim 12, the combination of Castiglione and Li teaches method according to claim 1, further comprising:
determining a plurality of fairness impact values corresponding to the plurality of original samples by changing the plurality of original samples in the original sample set respectively (Castiglione, [0175] – “Counterfactual Fairness (CFF), like FTA, mandates that similar individuals be treated similarly, but it does so causally. More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c.”, [0261], and in [0329] – “In various embodiments, the (input) value is one of a plurality of values configured to be received by the one or more processors, the fairness indicator value is one of a plurality of fairness indicator values, each of the plurality of fairness indicator values generated based on a corresponding one of the plurality of values, and the output data is indicative of whether an aggregated measure of the plurality of fairness indicators exceeds a predefined fairness threshold.” – teaches determining a plurality of fairness impact values (plurality of generalized local fairness values generated using plurality of fairness indicator values) corresponding to the plurality of original samples by changing the plurality of original samples in the original sample set respectively (generalized local fairness values correspond to plurality of original samples x by modifying the plurality of original samples in the original sample set));
generating an updated model based on the [[an]] updated sample set (Castiglione, [0036] – “In an alternate embodiment, the specialized computing system is utilized as an auditor or adjudicator subsystem that receives input models, and is configured to automatically flag or cause the models to be re-trained, re-engineered, or removed from service for an estimated fairness violation.” and in [0037] – “Re-training can include resetting or randomizing model weights and re-conducting training, training with different training sets, among others (e.g., re-training a model using training set B instead of training set A). Re-engineering can include changing the structures of the models or removing features or nodes from analysis from the models (e.g., changing the possible input feature sets or removing nodes from the latent space).” – teaches generating an updated model based on an updated sample set (computing system utilized as an auditor that automatically flags models to be re-trained or re-engineered. Re-training includes training with different sets, re-engineering includes changing model structure or removing features/nodes from analysis from the models)).
The combination of Castiglione and Li fails to explicitly teach determining, based on the plurality of fairness impact values, a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples.
However, analogous to the field of the claimed invention, Koh teaches:
determining, based on the plurality of fairness impact values, a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples (Koh, Section 2.1 Last Paragraph – “By setting δ in the direction of Ipert,loss(z,ztest), we can construct local perturbations of z that maximally increase the loss at ztest. In Section 5.2, we will use this to construct training-set attacks. Finally, we note that Ipert,loss(z,ztest) can help us identify the features of z that are most responsible for the prediction on ztest.” – teaches determining a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples based on the plurality of fairness impact values (utilizes influence functions to identify features with the largest impact on prediction, thus determining a plurality of target samples having largest corresponding fairness impact values from a plurality of original samples));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the determination of target samples of Koh to the samples, models, and plurality of fairness impact values of Castiglione and Li. Doing so would provide methods for identifying the training points most responsible for a given prediction and reveal insights about how to rely on and extrapolate from training date (Koh, Section 5.1).
The combination of Wang, Castiglione, and Koh fails to explicitly teach changing labels of the plurality of original samples in the original sample set respectively; and generating an updated sample set by changing the labels of the plurality of target samples.
However, analogous to the field of the claimed invention, Wang teaches:
changing labels of the plurality of original samples in the original sample set respectively (Wang, Section “Problem and Motivation” Paragraph 4 – “Specifically, for a sample (d,y), we get the corresponding counterfactual sample (d,y) by (i) substituting causal terms in d with their antonyms to get d’ and (ii) assigning an opposite label y’ to d’.” – teaches changing labels of the plurality of original samples in the original sample set (generates counterfactual samples with opposite label y’));
generating an updated sample set by changing the labels of the plurality of target samples (Wang, Section “Problem and Motivation” Paragraph 4 – “Specifically, for a sample (d,y), we get the corresponding counterfactual sample (d,y) by (i) substituting causal terms in d with their antonyms to get d’ and (ii) assigning an opposite label y’ to d’.” – teaches generating an updated sample set by changing labels of the plurality of target samples (generates counterfactual samples with opposite label y’));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the changing of labels of original samples of Wang to the original samples, target samples, models, and fairness impact values of Castiglione, Li, and Koh. Doing so would provide methods for automatic generation of counterfactual samples with opposite labels to train robust classifiers (Wang, Introduction).
Claim(s) 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Castiglione and Li as applied to claims 1, 15, and 20 above, and further in view of Koh and Dalli.
Regarding claim 13, the combination of Castiglione and Li teaches the method according to claim 1, further comprising:
determining a plurality of fairness impact values corresponding to the plurality of original samples by changing the plurality of original samples in the original sample set respectively (Castiglione, [0175] – “Counterfactual Fairness (CFF), like FTA, mandates that similar individuals be treated similarly, but it does so causally. More precisely, given a causal graphical model, a counterfactual from x may be generated by performing an intervention on the protected attribute c. This produces an x′ from the same value of z, but a different c.”, [0261, 0262], and in [0329] – “In various embodiments, the (input) value is one of a plurality of values configured to be received by the one or more processors, the fairness indicator value is one of a plurality of fairness indicator values, each of the plurality of fairness indicator values generated based on a corresponding one of the plurality of values, and the output data is indicative of whether an aggregated measure of the plurality of fairness indicators exceeds a predefined fairness threshold.” – teaches determining a plurality of fairness impact values (plurality of generalized local fairness values generated using plurality of fairness indicator values) corresponding to the plurality of original samples by changing the plurality of original samples in the original sample set respectively (generalized local fairness values correspond to plurality of original samples x by modifying the plurality of original samples in the original sample set));
generating an updated sample set by removing the [[a]] plurality of target samples from the original sample set (Castiglione, [0036] – “ In an alternate embodiment, the specialized computing system is utilized as an auditor or adjudicator subsystem that receives input models, and is configured to automatically flag or cause the models to be re-trained, re-engineered, or removed from service for an estimated fairness violation.” and in [0037] – “Re-training can include resetting or randomizing model weights and re-conducting training, training with different training sets, among others (e.g., re-training a model using training set B instead of training set A). Re-engineering can include changing the structures of the models or removing features or nodes from analysis from the models (e.g., changing the possible input feature sets or removing nodes from the latent space).” – teaches generating an updated sample set by removing a plurality of target samples from the original sample set (systems utilized as an auditor and automatically flags models to be re-engineered. Re-engineering includes removing features from analysis from the models, thus generating an updated sample set by removing a plurality of target samples from the original sample set)); and
generating an updated model based on the updated sample set (Castiglione, [0036] – “In an alternate embodiment, the specialized computing system is utilized as an auditor or adjudicator subsystem that receives input models, and is configured to automatically flag or cause the models to be re-trained, re-engineered, or removed from service for an estimated fairness violation.” and in [0037] – “Re-training can include resetting or randomizing model weights and re-conducting training, training with different training sets, among others (e.g., re-training a model using training set B instead of training set A). Re-engineering can include changing the structures of the models or removing features or nodes from analysis from the models (e.g., changing the possible input feature sets or removing nodes from the latent space).” – teaches generating an updated model based on the updated sample set (computing system utilized as an auditor that automatically flags models to be re-trained or re-engineered. Re-training includes training with different sets, re-engineering includes changing model structure or removing features/nodes from analysis from the models)).
The combination of Castiglione and Li fails to explicitly teach determining a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples based on the plurality of fairness impact values;
However, analogous to the field of the claimed invention, Koh teaches:
determining a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples based on the plurality of fairness impact values (Koh, Section 2.1 Last Paragraph – “By setting δ in the direction of Ipert,loss(z,ztest), we can construct local perturbations of z that maximally increase the loss at ztest. In Section 5.2, we will use this to construct training-set attacks. Finally, we note that Ipert,loss(z,ztest) can help us identify the features of z that are most responsible for the prediction on ztest.” – teaches determining a plurality of target samples having largest corresponding fairness impact values from the plurality of original samples based on the plurality of fairness impact values (utilizes influence functions to identify features with the largest impact on prediction, thus determining a plurality of target samples having largest corresponding fairness impact values from a plurality of original samples));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the determination of target samples of Koh to the samples, models, and plurality of fairness impact values of Castiglione and Li. Doing so would provide methods for identifying the training points most responsible for a given prediction and reveal insights about how to rely on and extrapolate from training date (Koh, Section 5.1).
The combination of Castiglione, Li, and Koh fails to explicitly teach changing values of specified features of the plurality of original samples in the original sample set respectively.
However, analogous to the field of the claimed invention, Dalli teaches:
changing values of specified features of the plurality of original samples in the original sample set respectively (Dalli, Pg. 15, Col. 6, Line 55 – Pg. 16, Col. 7, Line 15 – “Bias detection determines if a specific model or dataset is biased towards specific features. However, such features may depend on other features, which may result in a causal effect… Causal models may also be used in certain cases to correct or adjust for bias by creating what-if and what-if-not type of analyses and may also suggest suitable interventions. Causal models may also be used to detect the presence of bias using counterfactuals and may also determine how results from any of: (i.) one part of the model, (ii.) a related model, and/or (iii.) sub-part of the model, can be transferred or transposed directly or via the appropriate modification, conditioning, allowance or other appropriate transformation. Causal analysis may be used to enhance bias detection and mitigation with the application of mediation analysis, and sub-division of counterfactuals including direct and indirect effects of both interventional and counterfactual type of modifications.” and in Pg. 16, Col. 7, Line 65 – Pg. 17, Col. 8, Line 3 – “Bias detection may be detected by analysing the difference in feature attribution when applying a control swap. Control swap is a method whereby for a given transaction X, the potential data field which has a bias (for example gender), is modified to test any change in the outcome and explanation.” – teaches changing values of specified features of the plurality of original samples in the original sample set respectively (applies control swap to apply a difference in feature attribution of a sample, modifies data field of sample X change values of specified features)
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the feature modification of Dalli to the original samples, target samples, models, and fairness impact value determinations of Castiglione and Li. Doing so would provide methods for bias detection by analyzing differences in samples with modified features (Dalli, Pg. 16, Cols. 7-8).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Kamkar et al. (US Pub. No. 2022/0168477, published May 2022) teaches systems and methods for generating tree-based models with improved fairness. Teaches wherein fairness of a model is identified based on measured difference in outcomes for decisions based on the model for each value of a sensitive attribute. Teaches a custom loss function that receives original samples and samples with sensitive attributes. Teaches labeling samples with sensitive attributes.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
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, Matt Ell can be reached at 571-270-3264. 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.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights