TabClusterNet: Enhanced Deep Clustering for Tabular Data Analysis
<p>Shuwei Xu, Zhi Hu, Xiaowei Wang</p> · Academic Journal of Computing & Information Science · 2024
Clustering high-dimensional tabular data is a complex and challenging problem. Traditional clustering techniques nearly fail to identify the latent structure buried in spaces of high dimensionality. TabClusterNet is a novel deep clustering model designed specifically for tasks related to tabular data analysis. The self-supervised learning encoder-decoder from TabNet is combined with the deep clustering framework of Deep Embedding Clustering (DEC). By the high feature-extracting power of TabNet and the high clustering ability of DEC, TabClusterNet achieves far superior performance than the conventional method in feature extraction towards efficient clustering. Our proposed novel deep clustering architecture has been extensively validated over various public datasets for its great performance over different evaluation metrics. A closer look at the model shows that it preserves the structure of the data. TabClusterNet has been demonstrated to achieve substantially improved clustering accuracy and not only offer insights useful for data analytics and decision support, but also enable data scientists and researchers to glean deeper insights from complex datasets.