Information bottleneck disentanglement based sparse representation for fair classification

Abstract

Unlike current state-of-the-art methods based on data augmentation and adversarial frameworks to solve the fair classification problem, this paper proposes the Information Bottleneck Disentanglement Based Sparse Representation algorithm to ensure the fairness of extracted features. The algorithm assumes that each sample is collaboratively generated by a categorical latent variable and a sensitive or category-irrelevant latent variable. To ensure that these two latent variables correctly capture their corresponding features, we impose Information Bottleneck Disentanglement constraints on them. Specifically, the Information Bottleneck is used to obtain more relevant information while compressing the amount of information. Simultaneously, a mutual information minimization constraint enhances their independence. Additionally, by employing the sparse structure, we can further disentangle the discriminative and sensitive information under supervision guidance. Therefore, the algorithm can improve the discrimination of the categorical latent variable as much as possible without extracting the sensitive data into it. Extensive experiments on five public and challenging benchmark datasets reveal the effectiveness of our proposed method, demonstrating that it can ensure fairness without sacrificing classification performance.