Significance-based interpretable sequence clustering

Abstract

Recently, many interpretable clustering algorithms have been proposed, which focus on characterizing the clustering outcome in terms of explainable models such as trees and rules. However, existing solutions are mainly developed for handling standard vectorial data and how to obtain interpretable clustering results for complicated non-vector data such as sequences and graphs is still in the infant stage. In this paper, we present a significance-based interpretable clustering algorithm for discrete sequences, which has the following key features. Firstly, instead of using a third-party clustering method to obtain the initial clusters, we directly extract cluster-critical sequential patterns to describe potential clusters. Secondly, without needing to specify the number of clusters, we guide the growth of the decision tree through a hypothesis testing procedure. As a result, not only the final clustering result is explainable but also the tree construction process is statistically interpretable. Experimental results on real-world sequential data sets show that our algorithm achieves comparable performance to state-of-the-art methods in both cluster quality and interpretability.