Decision Tree Clustering for Time Series Data: An Approach for Enhanced Interpretability and Efficiency

Abstract

Clustering is one of the unsupervised learning methods for grouping similar data samples. While clustering has been used in a wide range, traditional clustering methods cannot provide clear interpretations of the resulting clusters. This has led to an increasing interest in interpretable clustering methods, which are mainly based on decision trees. However, the existing interpretable clustering methods are typically designed for tabular data and struggle when applied to time series data due to its complex nature. In this paper, we propose a novel interpretable time-series clustering method with decision trees. To address the interpretability challenges in time-series data, our method employs two separate feature sets, intuitive features for decision tree branching and original time-series observed values for evaluating a given clustering metric. This dual use enables us to construct interpretable clustering trees for time series data. In addition, to handle datasets with a large number of samples, we propose a new metric for evaluating clustering quality, called the surrogate silhouette coefficient, and present a heuristic algorithm for constructing a decision tree based on the metric. We show that the computational complexity for evaluating the proposed metric is much less than the silhouette coefficient, which is commonly used in decision tree-based clustering. Our numerical experiments demonstrated that our method constructed decision trees faster than the existing methods based on the silhouette coefficient while maintaining clustering quality. In addition, we applied our method to a time-series data on an e-commerce platform and succeeded in constructing an insightful decision tree.