TabM
Summary
TabM is a tabular deep-learning model that makes multiple predictions per object by packing a parameter-efficient ensemble of MLP-like submodels into one model.
Role In The Wiki
TabM is a strong static-tabular baseline, not a time-series model or tabular foundation model. It matters locally because it gives a practical, curated set of numerical feature embedding options for continuous columns: raw scalar inputs, simple linear-ReLU embeddings, piecewise-linear embeddings, and periodic embeddings.
Numerical Feature Interface
TabM’s numerical embeddings are typed feature embeddings. They are tied to table columns and can depend on feature-specific preprocessing or bins. This makes them a useful analogue for auxiliary numeric values in time-series models, such as exogenous variables, numeric control inputs, intervention intensities, and metadata, but they are not universal text-number tokens.
Official Artifacts
- Official code: yandex-research/tabm
- Numerical embedding package: yandex-research/rtdl-num-embeddings