Towards Learning Nonlinear Multivariate Correlations in Tabular Data

Learning complex dependencies in tabular data is a fundamental challenge in statistical learning. Traditional correlation measures often fail to capture nonlinear or multivariate relationships. In this paper, we introduce TMC, a neural framework for estimating maximal correlations across variables in tabular datasets. Our method supports pairwise (1 vs 1), multivariate (p vs 1), and groupwise correlation analyses. We evaluate TMC on both synthetic and real-world datasets, benchmarking it against existing correlation estimators. Results show that TMC consistently outperforms state-of-the-art methods by effectively capturing nonlinear dependencies and maintaining robustness under noise. Feature selection experiments further demonstrate its practical relevance.