Dynamic environment adaptive online learning with fairness awareness via dual disentanglement

Abstract

The widespread application of Artificial Intelligence (AI) comes with the necessity to consider and mitigate discrimination in machine learning algorithms. Most existing fair machine learning methods are only suitable for short-term and static scenarios, and thus cannot adapt to dynamically changing environments or meet the needs for real-time updates. In open dynamic scenarios, data arriving in batches needs processing in real-time, and the constantly changing environment will lead to data distribution shifts, making it difficult to ensure the fairness of models in the long run. To achieve long-term fairness of models, we propose an online dual disentanglement method that captures fair representations of non-sensitive core information in real-time within constantly changing environments, thereby enhancing the robustness of fair models. Firstly, learned representations are disentangled from environment-specific variation factors through a constrained optimization setup to ensure semantic invariance. Further, a bias disentanglement method based on supervised contrastive learning is designed. While keeping the non-sensitive core information unchanged, the sensitive information is hidden from semantic representations and the spurious correlation with target labels is cut off, so as to achieve the long-term fairness of the model decision. By formulating the fairness-aware online learning problem in dynamic environments as an online optimization problem with the long-term fairness constraint, and theoretically proving that the algorithm achieves sublinear dynamic regret and sublinear violation of cumulative unfairness under certain assumptions. Experimental evaluations on real-world datasets demonstrate the effectiveness of the proposed method, which maintains overall fairness above 80% without compromising utility, outperforming state-of-the-art baseline methods.