System and method for fairness aware optimization with features quantization
Various methods and processes, apparatuses or systems, and media for performing fairness aware optimization are disclosed. The present disclosure provides acquiring and quantizing a plurality of input features for an application, and binning of each of the plurality of input features quantized so that each of the plurality of input features is assigned to a bin among a plurality of bins to provide a matrix of bin membership for each of the plurality of applications. A vector is then generated based on the matrix of bin membership for evaluating the generated vector against a target value and presence of disparity in outcome for a protected input feature. Once a coefficient vector that optimizes an output while negating any disparity in outcome for the protected input feature is identified, an optimization model is updated with such coefficient vector for subsequent processing.
1 . A method for performing a fairness aware optimization on a plurality of attributes, the method comprising:
acquiring, by a processor and from a plurality of databases, an input dataset including a plurality of applications with corresponding set of input features, wherein each of the plurality of applications has a plurality of input features;
quantizing, by the processor, each of the plurality of input features for the plurality of applications;
performing binning, by the processor, of each of the plurality of input features quantized so that each of the plurality of input features for each of the plurality of applications is assigned to a bin among a plurality of bins to provide a matrix of bin membership for each of the plurality of applications;
querying, by the processor, a machine learning (ML) model for generating a vector for each of the plurality of applications based on a corresponding matrix of bin membership, wherein each of the plurality of bins is associated with a corresponding vector, such that each of the plurality of applications is characterized by a plurality of vectors based on the matrix of bin membership;
evaluating, via the ML model and for an application of the plurality of applications, the generated vector against a corresponding target value and presence of disparity in outcome for one or more of the plurality of input features, wherein the evaluating includes:
generating one or more function parameters for the corresponding matrix of bin membership;
generating a coefficient vector based on the one or more function parameters;
generating the corresponding target value based on the matrix of bin membership and the generated coefficient vector;
calculating a predicted output for the input dataset in view of the corresponding target value; and
calculating for the presence of disparity in outcome for at least one of the plurality of input features;
performing one or more training iterations of the evaluating, on the ML model and for remaining applications of the plurality of applications until an offset value for each bin that adjusts a regression target for input features grouped into a particular bin is determined;
identifying, by the ML model and based on performance of the one or more training iterations on the ML model, a target coefficient vector among a plurality of coefficient vectors generated for the plurality of applications that provides an optimized output while negating the presence of disparity in outcome;
storing one or more function parameters including the identified target coefficient vector, and updating the ML model with the target coefficient vector, wherein the stored one or more function parameters approximate an objective function of the ML model to limit search space for more efficient exploration with respect to unobserved points; and
applying the updated ML model on a subsequent application, wherein the updated ML model applies the offset at a bin level to provide an output without disparate outcome for any input feature of the subsequent application.
2 . The method according to claim 1 , wherein each of the plurality of bins are defined by a pair of boundary values.
3 . The method according to claim 1 , wherein the plurality of bins are of equal size.
4 . The method according to claim 1 , wherein the plurality of bins include bins of uneven sizes.
5 . The method according to claim 1 , wherein the ML model is a Bayesian optimization model.
6 . The method according to claim 1 , wherein each of the plurality of applications has its input features assigned to less than all of the plurality of bins.
7 . The method according to claim 1 , wherein the optimized output refers to a highest output value that negates the presence of disparity in outcome for a protected feature among the plurality of input features.
8 . The method according to claim 7 , wherein the protected feature is a feature regulated by a government entity.
9 . The method according to claim 7 , wherein the protected feature is a race of an applicant of at least one of the plurality of applications.
10 . The method according to claim 7 , wherein the protected feature is a gender of an applicant of at least one of the plurality of applications.
11 . A system for performing a fairness aware optimization on a plurality of attributes, the system comprising:
a processor; and
a memory operatively connected to the processor via a communication interface, the memory storing computer readable instructions, when executed, causes the processor to perform:
acquiring, from a plurality of databases, an input dataset including a plurality of applications with corresponding set of input features, wherein each of the plurality of applications has a plurality of input features;
quantizing each of the plurality of input features for the plurality of applications;
performing binning of each of the plurality of input features quantized so that each of the plurality of input features for each of the plurality of applications is assigned to a bin among a plurality of bins to provide a matrix of bin membership for each of the plurality of applications;
querying a machine learning (ML) model for generating a vector for each of the plurality of applications based on a corresponding matrix of bin membership, wherein each of the plurality of bins is associated with a corresponding vector, such that each of the plurality of applications is characterized by a plurality of vectors based on the matrix of bin membership;
evaluating, via the ML model and for an application of the plurality of applications, the generated vector against a corresponding target value and presence of disparity in outcome for one or more of the plurality of input features, wherein the evaluating includes:
generating one or more function parameters for the corresponding matrix of bin membership;
generating a coefficient vector based on the one or more function parameters;
generating the corresponding target value based on the matrix of bin membership and the generated coefficient vector;
calculating a predicted output for the input dataset in view of the corresponding target value; and
calculating for the presence of disparity in outcome for at least one of the plurality of input features;
performing one or more training iterations of the evaluating, on the ML model and for remaining applications of the plurality of applications until an offset value for each bin that adjusts a regression target for input features grouped into a particular bin is determined;
identifying, by the ML model and based on performance of the one or more training iterations on the ML model, a target coefficient vector among a plurality of coefficient vectors generated for the plurality of applications that provides an optimized output while negating the presence of disparity in outcome;
storing one or more function parameters including the identified target coefficient vector, and updating the ML model with the target coefficient vector, wherein the stored one or more function parameters approximate an objective function of the ML model to limit search space for more efficient exploration with respect to unobserved points; and
applying the updated ML model on a subsequent application, wherein the updated ML model applies the offset at a bin level to provide an output without disparate outcome for any input feature of the subsequent application.
12 . The system according to claim 11 , wherein each of the plurality of bins are defined by a pair of boundary values.
13 . The system according to claim 11 , wherein the plurality of bins are of equal size.
14 . The system according to claim 11 , wherein the plurality of bins include bins of uneven sizes.
15 . The system according to claim 11 , wherein the ML model is a Bayesian optimization model.
16 . The system according to claim 11 , wherein the optimized output refers to a highest output value that negates the presence of disparity in outcome for a protected feature among the plurality of input features.
17 . The system according to claim 16 , wherein the protected feature is a feature regulated by a government entity.
18 . The system according to claim 16 , wherein the protected feature is a race of an applicant of at least one of the plurality of applications.
19 . The system according to claim 16 , wherein the protected feature is a gender of an applicant of at least one of the plurality of applications.
20 . A non-transitory computer readable medium configured to store instructions for performing a fairness aware optimization on a plurality of attributes, when executed, cause a processor to perform:
acquiring, from a plurality of databases, an input dataset including a plurality of applications with corresponding set of input features, wherein each of the plurality of applications has a plurality of input features;
quantizing each of the plurality of input features for the plurality of applications;
performing binning of each of the plurality of input features quantized so that each of the plurality of input features for each of the plurality of applications is assigned to a bin among a plurality of bins to provide a matrix of bin membership for each of the plurality of applications;
querying a machine learning (ML) model for generating a vector for each of the plurality of applications based on a corresponding matrix of bin membership, wherein each of the plurality of bins is associated with a corresponding vector, such that each of the plurality of applications is characterized by a plurality of vectors based on the matrix of bin membership;
evaluating, via the ML model and for an application of the plurality of applications, the generated vector against a corresponding target value and presence of disparity in outcome for one or more of the plurality of input features, wherein the evaluating includes:
generating one or more function parameters for the corresponding matrix of bin membership;
generating a coefficient vector based on the one or more function parameters;
generating the corresponding target value based on the matrix of bin membership and the generated coefficient vector;
calculating a predicted output for the input dataset in view of the corresponding target value; and
calculating for the presence of disparity in outcome for at least one of the plurality of input features;
performing one or more training iterations of the evaluating, on the ML model and for remaining applications of the plurality of applications until an offset value for each bin that adjusts a regression target for input features grouped into a particular bin is determined;
identifying, by the ML model and based on performance of the one or more training iterations on the ML model, a target coefficient vector among a plurality of coefficient vectors generated for the plurality of applications that provides an optimized output while negating the presence of disparity in outcome;
storing one or more function parameters including the identified target coefficient vector, and updating the ML model with the target coefficient vector, wherein the stored one or more function parameters approximate an objective function of the ML model to limit search space for more efficient exploration with respect to unobserved points; and
applying the updated ML model on a subsequent application, wherein the updated ML model applies the offset at a bin level to provide an output without disparate outcome for any input feature of the subsequent application.