Why do tree-based models outperform deep learning on tabular data?

“The man who knows how will always have a job. The man who knows why will always be his boss.” – Ralph Waldo Emerson

The study shows that tree-based methods consistently outperform neural networks on tabular data with about 10K observations, both in prediction error and computational efficiency, with and without hyperparameter tuning. 45 datasets from different domains are modeled for benchmarking.

The paper then goes on to explain why. The “why” part offers some experiments but looks quite empirically driven so I can’t say I’m convinced there. The Hugging Face repo for the paper, datasets, code, and a detailed description is a great resource though.

Source