Due to hard accessibility, real-world adoption of fair representation learning algorithms lacks the prior knowledge of the sensitive attributes that we wish to be fair with. To address the challenge in fairness without explicit demographics, our solution is based on the idea of maximally randomizing the representation while being as informative as possible about the target task. We operationalize this goal through the concept of maximizing the entropy of the learned representation. For this purpose, we propose two new avenues for entropy maximization in the absence of demographic information: intra-class and inter-class entropy maximization. For 1) intra-class entropy maximization, it maximizes the entropy of the non-target class predictions (excluding the probability of the ground truth class label for classification problems), thus encouraging the model to discard spurious correlations between the different target classes, and for 2) inter-class entropy maximization, it maximizes the entropy of the representation conditioned on the target label, thus encouraging randomization of the samples within each target class label and minimizing the leakage of potential demographic information in the representation. Quantitative and qualitative results of our Maximum Entropy method (MaxEnt) on COMPAS and UCI Adult datasets show that 1) our method can outperform the State-of-the-art (SOTA) Adversarially Reweighted Learning (ARL) method and will enhance the difficulty of extracting sensitive demographic information in representation without prior demographic knowledge 2) our method reaches a good trade-off between utility and fairness.
Read
