Deep Learning Ensemble Improves Imbalanced Tabular Data Clustering

Data imbalance is a significant hurdle in supervised classification, often leading to biased models that favor majority classes and produce misleading accuracy metrics. Unsupervised deep clustering, however, offers a potential solution as its representation learning process does not rely on class labels, making it inherently more robust to such imbalances. While deep clustering has been widely applied to image, language, and graph data, its use for tabular data is a more recent development. This research systematically evaluates state-of-the-art deep clustering methods under varying degrees of data imbalance. It proposes two innovative ensemble approaches: one that combines deep clustering assignments across different embedding dimensions, and another that uses majority voting among the best-performing clustering algorithms. Experiments on 16 diverse binary tabular datasets, with both natural and artificially induced imbalance, demonstrate that these ensemble methods consistently achieve higher accuracy, normalized mutual information (NMI), and adjusted Rand index (ARI) scores compared to individual clustering methods. This makes deep clustering a strong alternative to supervised classification in scenarios with imbalanced tabular data.