Leveraging Structural Causal Models for Bias Detection and Feature Adjustment in Machine Learning Predictions
Bern Igoche Igoche, Olumuyiwa Matthew, Peter Bednár, ALEXANDER E. GEGOV · 2024
This research addresses the critical issue of fairness bias in machine learning predictions and explanations around the Nigerian Polytechnic admission process. We propose a novel approach that leverages Structural Causal Models (SCMs) to design an ontological framework for detecting bias and adjusting features in Local Interpretable Model-agnostic Explanations (LIME). The SCM ontology provides a principled approach to identify the features that contribute to fairness bias, and we show the presence of bias in LIME explanations for the Polytechnic admission dataset. To mitigate this bias, we propose an ablation technique for feature adjustment in a fair-LIME framework that uses SCM ontology. Experimental results on the Benpoly admission dataset show a remarkable reduction in fairness bias with high fidelity to the predictions of the black-box model. The fair-LIME explanations provided a valid and unbiased interpretation of the factors driving admission decisions. In this regard, the paper significantly extends the scope of eXplainable AI by developing a principled method for obtaining an unbiased explanability aimed at fostering the growth of trustworthy AI systems. The fair-LIME framework holds strong implications towards making the development of transparent, accountable, and ethical AI systems a reality for high-stakes decision-making processes.