Interpretable Clustering via Soft Clustering Trees

Clustering is a popular unsupervised learning task that consists of finding a partition of the data points that groups similar points together. Despite its popularity, most state-of-the-art algorithms do not provide any explanation of the obtained partition, making it hard to interpret. In recent years, several works have considered using decision trees to construct clusters that are inherently interpretable. However, these approaches do not scale to large datasets, do not account for uncertainty in results, and do not support advanced clustering objectives such as spectral clustering. In this work, we present soft clustering trees, an interpretable clustering approach that is based on soft decision trees that provide probabilistic cluster membership. We model soft clustering trees as continuous optimization problem that is amenable to efficient optimization techniques. Our approach is designed to output highly sparse decision trees to increase interpretability and to support tree-based spectral clustering. Extensive experiments show that our approach can produce clustering trees of significantly higher quality compared to the state-of-the-art and scale to large datasets.