TabGSL: Graph Structure Learning for Tabular Data Predictions

Jay Chiehen Liao, Jun-Wei Chiu, Cheng-Te Li

Abstract: Graph structure learning (GSL) has shown promise in various domains, but its potential impact on tabular data prediction remains underexplored. Traditional deep learning models for tabular data often overlook implicit relationships between instances, which could provide valuable insights for improving predictive performance. While graph-based approaches have been successfully applied in domains with well-defined relational structures, tabular data lacks explicit graph connectivity, making it challenging to leverage graph neural networks (GNNs) effectively. This perspective paper presents a novel framework, Tabular Graph Structure Learning (TabGSL), to explore whether GSL can improve tabular data predictions. TabGSL unifies feature extraction with dynamic graph construction and contrastive learning, seamlessly integrating instance correlations and feature-level interactions within a single end-to-end architecture. Experimental evaluations on benchmark datasets reveal consistent improvements over both tree-based and neural baselines, suggesting that graph-induced relational insights drive better predictive performance and a more interpretable feature space, offering a new paradigm for tabular data prediction.

External IDs:dblp:conf/bigdataconf/LiaoCL25