The author discusses the challenge of working with data in tabular format and the recent use of Graph Neural Networks (GNNs) to handle this type of data effectively. However, GNNs often produce black-box models that make it difficult for users to understand the logic behind the predictions. To address this issue, the author proposes a new approach called IGNNet, which is an Interpretable Graph Neural Network for tabular data. IGNNet is shown to perform as well as other state-of-the-art machine learning algorithms for tabular data, such as XGBoost, Random Forests, and TabNet. Additionally, IGNNet provides explanations for its predictions that align with the true values of the features, without any extra computational burden.
https://arxiv.org/abs/2308.08945