Top-down induction of similarity measures using similarity clouds

Abstract

The automatic acquisition of a similarity measure for a CBR system is appealing as it frees the system designer from the tedious task of defining it manually. However, acquiring similarity measures with some machine learning approach typically results in some black box representation of similarity whose magic-like combination of high precision and low explainability may decrease a human user’s trust in the system. In this paper, we target this problem by suggesting a method to induce a humanreadable and easily understandable – and thus potentially trustworthy – representation of similarity from a previously learned black box-like representation of similarity measures. Our experimental evaluations support the claim that, given some highly precise learned similarity measure, we can induce a less powerful, but human-understandable representation of it while its corresponding level of accuracy is only marginally impaired.