Abstract

Deep learning has demonstrated remarkable success across domains such as image, audio, video, and text processing. However, its application to tabular data remains challenging due to the heterogeneous nature of features and the lack of positional information, which complicates feature interaction modeling. While deep learning approaches for tabular data have achieved performance levels comparable to traditional methods like gradient boosting machines, their adoption in real-world scenarios is limited by issues of interpretability and robustness. We introduce FusANet (Fusion Attention Neural Network), a novel parametric transformer-style architecture for tabular data that fuses a dataset-level global attention map with instance-level local self-attention using a cross-attention fusion module. FusANet (i) learns a shared, pre-initialized global attention matrix W_global to capture population-wide feature relationships, (ii) retains sample-specific interactions via local multi-head self-attention, and (iii) fuses these views with cross-attention to produce robust predictions and dual-level explanations. Across 10 benchmark tabular datasets, we show FusANet achieves average improvements up to 19.72 % in classification accuracy and up to 24.25 % relative RMSE reduction on regression task versus strong parametric state-of-the-art models and performs competitively with gradient-boosted trees, while offering dataset- and instance-level attributions.