Tree-regularized tabular embeddings improve performance and robustness of neural networks for tabular data.