Tabular Representation Learning

Tabular representation learning studies how to encode rows, columns, and whole tables so models can use mixed numerical and categorical data, missing values, and schema structure. Common approaches include feature-aware transformers, masked or contrastive self-supervised learning, autoencoders, and embeddings that combine table structure with language-model or graph representations. These embeddings support tasks such as classification, retrieval, clustering, anomaly detection, synthetic data generation, and transfer across datasets or schemas. Results remain sensitive to the task and dataset: no single encoder consistently leads, and when pretraining helps over training from scratch or tree-based models is still uncertain.