Neural Networks Bias Mitigation Through Fuzzy Logic and Saliency Maps
Sahar Shah, Davide Ciucci, Sara Manzoni, Italo Francesco Zoppis · 2025
Mitigating biases in neural networks is crucial to reduce or eliminate the predictive model’s unfair responses, which may arise from unbalanced training, defective architectures, or even social prejudices embedded in the data. This study proposes a novel and fully differentiable framework for mitigating neural network bias using Saliency Maps and Fuzzy Logic. We focus our analysis on a simulation study for recommendation systems, where neural networks are crucial in classifying job applicants based on relevant and sensitive attributes. Leveraging the interpretability of a set of Fuzzy implications and the importance of features attributed by Saliency Maps, our approach penalizes models when they overly rely on biased predictions during training. In this way, we ensure that bias mitigation occurs within the gradient-based optimization process, allowing efficient model training and evaluation.