Training machine learning models based on class-aware bias mitigation

Abstract

Techniques relate to detecting and reducing bias in regression-based machine learning models. A baseline model is trained using a regression algorithm and a baseline training data set comprising examples with independent parameters and corresponding continuous dependent values. A classification algorithm is executed on the training data set to determine, for at least a portion of examples, a probability distribution of class assignment across a set of classes. Class-aware machine learning models are trained using examples in accordance with their probability of belonging to respective classes. For each example, outputs from the baseline and class-aware models are combined according to class assignment to determine an expected class-aware output. A difference between the expected class-aware output and the baseline model output is evaluated. If bias is detected based on a predefined threshold, a parameter correlated with the differences is identified and excluded, and the baseline model is retrained with the modified data set.