Fuzzy-Vocabulary-Based Detection and Explanation of Anomalies

Abstract

Fuzzy partitions associated with linguistic variables are particularly useful to provide users with a description of the data. However, designing a fuzzy vocabulary that makes it possible to linguistically describe the data distribution and its inner structure is a tedious task. This paper introduces a novel strategy to infer possible fuzzy partitions from the data distribution with the objective to have available modalities to describe both dense and sparse regions. A data inner structure as well as the anomalies are then identified using this vocabulary whose terms are also used to provide users with contrastive explanations about the found anomalies.