In-Processing Neuro-Fuzzy Approaches for Fairness-Aware and Explainable Machine Learning

Machine learning (ML) now plays a direct role in decisions that shape people’s access to healthcare, credit, employment, and justice. When these systems learn from data that carries social bias, they can unintentionally reinforce inequality rather than help reduce it. This thesis focuses on improving fairness through in-processing strategies while keeping decision processes understandable to those affected by them. The central idea is to blend the transparency of fuzzy rules with the predictive strengths of neural networks (NNs). In doing so, the work explores how sensitive information influences model behaviour and proposes mechanisms to keep that influence under control. The thesis unfolds in three main stages. The first contribution introduces an in-processing learning strategy that uses attribution signals to identify when sensitive features dominate predictions and suppresses such behaviour during training. The second contribution updates the Adaptive Neuro-Fuzzy Inference System (ANFIS) by integrating kernel functions to increase flexibility and by injecting attribution feedback to improve stability and rule usefulness. The third contribution builds on this idea by allowing both the rule base and membership functions to evolve while the network trains, making the system more responsive to how bias appears across different data contexts. These developments are later unified into a single framework that works with commonly used attribution techniques, such as Integrated Gradients (IGs) and SmoothGrad. The framework reports fairness indicators alongside classification performance and offers interpretable, rule-level insights into where harmful bias may arise. The methods are evaluated on simulated datasets and public benchmarks, including Adult, Correctional Offender Management Profiling for alternative sanctions (COMPAS), and German Credit. Across all cases, the proposed models reduce group-level disparity measures such as statistical parity difference and disparate impact, while maintaining accuracy comparable to standard baselines. The results confirm that it is possible to improve both interpretability and fairness at the same time when neuro-fuzzy reasoning is embedded into the core of the learning process.