Kernelizing Adaptive Neuro-Fuzzy Inference for Bias Mitigation

Adaptive Network-based Fuzzy Inference System (ANFIS) is a powerful hybrid model that combines neural networks and fuzzy logic, making them highly effective for complex decision making. However, their use in sensitive contexts raises critical and urgent concerns about bias and fairness, as biased predictions can inadvertently embed systemic discrimination and inequality. In this study, we improved the capabilities of a traditional ANFIS to address the challenge of bias mitigation. We first defined specific rules based on the saliency (importance) of input features to quantify and reduce bias through system regularization. Then, to further strengthen the system capabilities, we kernelized the system logic, adding nonlinearity to the decision-making process.Experimental results, based on simulated data, demonstrate the effectiveness of our approach in reducing saliencies of sensitive features in the system output.