Fast and Accurate Explanations of Distance-Based Classifiers by Uncovering Latent Explanatory Structures

Distance-based classifiers like k-nearest neighbors and support vector machines are widely used in real-world applications, but explaining their predictions has been challenging—this paper addresses that gap by developing new methods to make these common models more transparent to users.

The researchers discovered that distance-based classifiers contain hidden neural network-like structures that can be leveraged for explanations, allowing established explainability techniques (specifically layer-wise relevance propagation) to be applied to models where they weren't previously usable.

The proposed explanation method outperformed several baseline approaches in quantitative evaluations, suggesting it provides more reliable and accurate insights into why these classifiers make specific predictions compared to existing alternatives.

The practical demonstration through two real-world use cases shows that these explanations have genuine utility beyond theoretical interest, indicating the approach could help practitioners understand and trust distance-based models in actual deployment scenarios.

This work bridges a gap in responsible AI by making a class of widely-deployed models more interpretable, which is important for building user trust and enabling human oversight in applications where model decisions affect real outcomes.