Counterfactual Data Augmentation for Predictive Models

This technical report presents a novel method for counterfactual data augmentation aimed at enhancing the performance of deep learning predictive models. The report outlines the limitations of traditional methods for addressing biases in training data, which often lead to over-reliance on spurious correlations. It introduces counterfactual inference as a technique for synthesizing plausible images by modifying specific attributes while preserving others, thereby eliminating false associations. The methodology includes two main approaches: expanding datasets through counterfactual data augmentation and adjusting the training objective with a counterfactual regularization term. The report also compares the effectiveness of the counterfactual method against commonly used debiasing techniques, focusing on both local performance within predefined attribute subgroups and overall global performance. Additionally, it evaluates the ability of trained image classifiers to adapt to counterfactual examples and assesses their fairness concerning selected attributes, contributing to the understanding of bias mitigation in machine learning applications.