Tabular Imbalanced Learning: A Survey, Benchmark, and Practical Guide

Abstract

Imbalanced learning remains a fundamental challenge in tabular data applications. Despite decades of research and numerous proposed methods, there is still limited systematic understanding of how different imbalance-handling strategies perform across diverse data regimes and computational constraints, making practical method selection difficult. In this work, we provide a systematic survey of tabular imbalanced learning and introduce Tabular Imbalanced Learning Benchmark (TILBench), a large-scale empirical benchmark for evaluating existing methods. We first organize imbalanced learning approaches into a unified taxonomy and then benchmark more than 40 representative methods across 57 tabular datasets under a standardized evaluation protocol, examining overall predictive performance, sensitivity to dataset characteristics, and computational scalability. Our results show that no single method consistently dominates across all settings. Instead, the effectiveness of different strategies depends strongly on dataset regimes and computational constraints, highlighting the need for regime-aware method selection. Based on these findings, we provide practical recommendations for selecting imbalanced learning methods under different data conditions and identify open challenges and promising directions for future research.