Visualizing Pairwise Feature Interactions in Neural Additive Models

Abstract

We present an approach for incorporating feature interactions into Neural Additive Models (NAMs), building upon existing work in this area, to enhance their predictive capabilities while maintaining interpretability. Our contribution focuses on the visual exploration and management of the increased number of feature maps resulting from the addition of pairwise feature combinations to NAMs. This method allows for effectively visualizing individual and pairwise feature interactions using line plots and heatmaps, respectively. To address the potential explosion in the number of feature maps, we apply different scoring functions to compute the importance of a feature map and then filter and sort them based on their importance. The proposed interactive dashboard effectively manages large sets of feature maps, while preserving the white-box properties of NAMs.