Go to ICLR 2025 Conference homepageMaskTab: Masked Tabular Data Modeling for Learning with Missing Features
Tabular machine learning has garnered increasing attention due to its practical value. Unlike the complete and standardized data often assumed in academia, tabular data primarily originates from industrial contexts and usually faces the issue of incomplete data samples, i.e., some features of a sample may be unpredictably missing. In this work, we introduce MaskTab, a masked tabular data modeling framework designed to facilitate model learning despite missing features. Instead of pursuing to accurately restore missing features like existing imputation methods, we jointly approach missing feature modeling and downstream tasks (e.g., classification) with a unified objective. Concretely, we propose to randomly drop out some solid features during training, equipped with a missing-related masked attention mechanism, to help the model rely more on trustworthy features when making decisions. Experiments on the very recent industry-grade benchmark, TabReD, suggest that our method surpasses the second DNN-based competitor by a clear margin, demonstrating its effectiveness and robustness in real-world scenarios. We will release the code and the model to facilitate reproduction.