Title: Generalizing Fair Null-Space Projections for Continuous Attributes to Kernel Methods

Abstract:

As machine learning systems become increasingly embedded in the daily social lives of millions, ensuring fairness has emerged as a critical priority in their development. Standard fairness frameworks typically depend on protected attributes to identify potential biases. However, existing research predominantly addresses discrete scenarios for both target and protected variables. There is a notable scarcity of literature concerning continuous attributes, particularly within the context of regression—a domain we define as continuous fairness.

Current strategies often employ iterative null-space projection; however, this technique has thus far been limited to linear models or embeddings derived from non-linear encoders. In this work, we advance this approach by extending null-space projections into kernel-induced feature spaces utilizing the "empirical feature space." We provide a theoretical derivation of this method as a direct transformation of the kernel matrix, resulting in a model- and fairness-score-agnostic technique suitable for continuous protected attributes. Our experiments demonstrate that integrating this novel approach with Support Vector Regression (SVR) yields competitive or superior performance across various datasets when compared to other contemporary methods.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC