An End-to-End Framework for Fair and Transparent Decision Support using Explainable Deep Learning

Abstract

This paper introduces FairXDL, a decision support system that jointly maximizes prediction accuracy and demographic fairness via a differentiable fairness regularizer. Leveraging the power of a compact residual deep neural network architecture, FairXDL uses Integrated Gradients to produce human-readable interpretations of individual samples without any additional forward passes. In experiments on the COMPAS recidivism dataset ($n = 7,214$), FairXDL increases the demographic parity ratio by 18.4 percentage points while suffering no more than a 0.9% reduction in prediction accuracy compared to a fairness-naive model. A streamlined ONNX export pipeline verifies the feasibility of deploying this approach in production.