Edge-updating graph neural networks for modeling feature interactions in tabular data

We proposed a message-passing graph neural network (GNN) based on graph isomorphism network (GIN) for learning on tabular data. Fully-connected, unweighted feature graphs were constructed from tabular data using contextual feature encodings for numerical and categorical features. A classification node was added to the feature graph to represent the entire graph during inference. Feature interactions were modeled using the proposed architecture, which used a neural network to learn edge attributes while leveraging residual connections for both node and edge updates to alleviate oversmoothing, a commonly-faced problem in GNNs. Our model was evaluated on 12 publicly available datasets and achieved the best mean rank among 6 tabular deep learning and GNN models. Furthermore, since gradient-boosted decision trees are considered to be state-of-the-art models for tabular data, we compared our model to XGBoost and CatBoost and found that our model outperformed both models in all datasets when using default hyperparameters and achieved the best results in 8 datasets when using tuned hyperparameters. A further comparison was made against 5 commonly used or recently-proposed GNNs to investigate the effectiveness of our model, in which our model achieved the top result in all datasets.