AM-DEC: Attention-based Mixed-type Deep Embedded Clustering

Mixed-type data, combining numerical and categorical attributes, is prevalent in domains such as healthcare, finance, and social science, yet poses challenges for traditional clustering algorithms. While deep learning based clustering excels on numerical datasets, few methods address heterogeneous data. This paper proposes a novel deep clustering framework AM-DEC, that learns a unified embedding space for mixed-type data via Variational Auto-Encoder by jointly encoding numerical and categorical features, with a Self-Attention mechanism to capture cross-feature correlations. We analyzed and compared the strengths and limitations of other works by presenting a systematic review of clustering techniques for mixed-type data. Our experiment results demonstrate that AM-DEC outperforms state-of-the-art baselines on various real-world datasets in Rand Index score and Mutual Information Index, offering a new way for heterogeneous data analysis.