Evaluating Prototype Explanations in Machine Learning

Prototype-based post-hoc explanations aim to make model predictions interpretable by presenting representative examples (prototypes) that illustrate how the model arrives at decisions. Their evaluation often relies on quantitative metrics such as fidelity (how closely prototypes approximate the model’s decision function), coverage (how much of the input space they represent), stability (whether explanations remain consistent under small perturbations), and diversity (ensuring prototypes capture distinct aspects of the data). In practice, however, these metrics face challenges: label noise or biased training data can cause prototypes to reflect spurious correlations rather than meaningful patterns, meaning high fidelity to a noisy model may reduce their reliability. This issue is particularly critical in high-stakes domains such as healthcare, anomaly detection, and natural language processing, where biases in text corpora may cause prototypes to reinforce stereotypes rather than clarify model reasoning.