Optimizing Arithmetic Feature Interaction for Enhanced Deep Tabular Learning

With the wide application of deep learning in tabular data processing, how to effectively capture the interaction between features has become one of the key factors to improve model performance. The academic community has proposed the necessity of arithmetic feature interaction to deal with low inductive bias in deep table learning. However, this method still has some limitations when dealing with complex feature interactions. Based on the research, this paper makes improvements to further improve the performance of deep learning models on tabular data and proposes two optimized arithmetic feature interaction strategies: 1) Dynamic Gated Weight Allocation (DGW); 2) Hierarchical Progressive Interaction Network (HPIN). The superiority of these methods on multiple standard data sets are verified by experiments. The experimental results show that the proposed methods significantly improve the accuracy and generalization ability of the model, especially on the dataset with high dimensions and complex relationships. Our research provides new ideas for deep table learning and shows the potential of interactive optimization of arithmetic features in practical applications.