Interpretable graph neural networks for tabular data

arxiv.org

77 points·PaulHoule··5 comments

> At the same time, the results show that the explanations obtained from IGNNet are aligned with the true Shapley values of the features without incurring any additional computational overhead

TabPFN: https://github.com/automl/TabPFN https://twitter.com/FrankRHutter/status/1583410845307977733

"TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second" (2022) https://arxiv.org/abs/2308.08945

FWIU TabPFN is Bayesian-calibrated/trained with better performance than xgboost for non-categorical data

Right, the significance of the original article and the related field of research is that ChatGPT-like models don't handle tabular data well and there's a lot of need for things that do.

There are multiple metrics to optimize for when optimizing.

FWIU, from the diagram in the photo in the linked tweet, which is similar to a diagram on page 16 of the TabPFN paper [1], on the OpenML-CC18, TabPFN has a better ROC Receiver Operating Characteristic after 1 second than XGboost, Catboost, LightGBM, KNN, SAINT, Reg. Cocktail, and Autogluon after any amount of time, but Auto-sklearn 2.0 required 5 minutes to reach ~ROC parity with TabPFN.

1. "TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second" (May 2023) https://arxiv.org/abs/2207.01848

2. "Interpretable Graph Neural Networks for Tabular Data" (Aug 2023) https://arxiv.org/abs/2308.08945

Microsoft LIDA seems something that can incorporate this for additional insight from tabular data sources

The paper sounds interesting, but without the code being available, it is difficult to say whether the idea is really viable.

Reply on news.ycombinator.com