An Evaluation Methodology for Explanations of Clustering Results in Textual Domain

Abstract

Explainability has become a must for any AI algorithm acting as a black-box, like Graph Spectral Clustering (GSC). Though a success was achieved in developing respective explanation methodologies, a new challenge has to be faced: the evaluation of the (quality of) explanations. Several evaluation methods for explanations in AI have been developed, but they turn out to have some shortcomings, making them unsuitable for GSC explanation evaluation. In this paper, we recall some of these methods and point out their respective shortcoming. Based on this investigation, we suggest a new explanation methodology oriented towards GSC and similar methods and present a small experiment on the usage of this methodology.