Neural Additive Model

The core idea of SIAN is the Feature Interaction Selection (FIS) algorithm: it first trains a reference DNN, then applies Archipelago-based interaction detection to measure the interaction strength for each feature subset. Using a heredity condition (threshold $\tau$), it explores higher-order interaction candidates only when lower-order interactions are significant, thereby avoiding an exhaustive search. Based on the selected interaction set , it constructs and trains a GAM-style additive model .