Interpretable Clustering Using Dempster-Shafer Theory

Abstract

This study presents DSClustering, a novel algorithm that merges clustering validity with interpretability using the Dempster-Shafer theory. Traditional clustering methods like Kmeans, DBSCAN, and agglomerative clustering, while widely used for their efficiency and accuracy, often fall short in transparency, creating barriers in critical fields such as healthcare, finance, and consumer analytics where decision-making requires clear, interpretable insights. DSClustering aims to bridge this gap by assigning clusters based on belief functions from Dempster-Shafer theory, which allows it to generate rule-based explanations for each data point’s cluster assignment. Through detailed experiments on real-world datasets, including consumer behavior and airline satisfaction data, we evaluate DSClustering against traditional algorithms using key metrics such as Silhouette score, Rand index and Dunn’s index for clustering validity. The results indicate that DSClustering not only performs competitively but also offers a clear interpretative layer, making it suitable for applications where understanding model outputs is as essential as the accuracy of the outputs themselves. This work underscores the increasing importance of interpretability in machine learning, particularly in unsupervised learning, where transparency is typically challenging to achieve. DSClustering demonstrates a promising approach for balancing robust clustering with user-oriented interpretability, potentially encouraging broader adoption of interpretable clustering models in data-critical industries.