Sparse clustering of mixed data with likelihood based feature ranking

Abstract

Mixed data refer to tabular data which include both numerical and categorical features and they have become prevalent in various fields such as finance and medical studies. In our study, we propose a simple yet powerful sparse clustering technique for mixed data. Our approach combines model-based Gaussian-multinomial mixture model with partitioning method, leveraging the advantages of both. Also, we utilize the difference in log-likelihoods between cluster assignment and non-assignment of each feature to induce sparsity and feature selection tailored to the practitioner's needs. The proposed method performs under with high-dimensional settings where the number of features exceeds the number of observations, due to its straightforward structure and capacity to induce sparsity. Furthermore, our model can select different features for each cluster and offers feature importance rankings which greatly enhances interpretability of the clustering result compared to other sparse clustering techniques for mixed data. We demonstrate our method's performance using synthetic and real data and observe that it has competitive performance compared to some state-of-the-art mixed data clustering methods.