Systems and methods for machine learning using adversarial learning and data transformation

A method for detecting credit card fraud in a computing environment, comprising:

• obtaining transaction data associated with credit card usage, wherein the transaction data includes a plurality of transaction features and a fraud indicator for each transaction;
• preprocessing the transaction data to transform the plurality of transaction features into dense continuous-type numeric vectors suitable for logistic regression analysis;
• augmenting the transaction data to address data imbalance between fraudulent and non-fraudulent transactions by employing a Negative-Calibrated Generative Adversarial Network (NCGAN) that generates synthetic fraudulent transaction data distinct from non-fraudulent transaction data;
• modeling adversarial behavior of fraudsters by adjusting features of fraudulent transactions in the augmented transaction data based, at least in part, on parameters of a logistic regression model, wherein the adjustment simulates a response to the logistic regression model in accordance with a simulated Stackelberg game;
• training the logistic regression model using a sum of top-K loss function that prioritizes a subset of transaction data associated with a greatest loss to refine the parameters of the logistic regression model;
• and
• classifying transactions as fraudulent or non-fraudulent in a test dataset based, at least in part, on the refined parameters of the logistic regression model.

A non-transitory computer readable medium containing computer executable instructions that, when executed by a computer processor, cause the computer processor to perform operations, comprising:

• obtaining transaction data associated with credit card usage, wherein the transaction data includes a plurality of transaction features and a fraud indicator for each transaction;
• preprocessing the transaction data to transform the plurality of transaction features into dense continuous-type numeric vectors suitable for logistic regression analysis;
• augmenting the transaction data to address data imbalance between fraudulent and non-fraudulent transactions by employing a Negative-Calibrated Generative Adversarial Network (NCGAN) that generates synthetic fraudulent transaction data distinct from non-fraudulent transaction data;
• modeling adversarial behavior of fraudsters by adjusting features of fraudulent transactions in the augmented transaction data based, at least in part, on parameters of a logistic regression model, wherein the adjustment simulates a response to the logistic regression model in accordance with a simulated Stackelberg game;
• training the logistic regression model using a sum of top-K loss function that prioritizes a subset of the transaction data associated with a greatest loss to refine the parameters of the logistic regression model;
• and
• transactions as fraudulent or non-fraudulent in a test dataset based, at least in part, on the refined parameters of the logistic regression model.