[4I3-OS-1b] AutoML（自動機械学習）

[4I3-OS-1b-04] Neural Additive Models with Feature Selection and Pairwise Interactions

〇Yasutoshi Kishimoto¹, Kota Yamanishi¹, Takuya Matsuda¹, Shinichi Shirakawa¹ (1. Yokohama National University)

Keywords:Feature Selection, Interpretability, Neural Additive Models

Deep neural networks (DNNs) have been applied to various fields, but DNNs often face the problem of low interpretability. Neural Additive Models (NAM), which use DNN as shape functions in generalized additive models, are highly interpretable while retaining the flexibility of DNN training. NAM can provide and visualize the contribution of each feature to the prediction because one-input DNNs are used for transforming each feature. However, in the case of using two-input DNNs to take into account feature interactions or applying NAM to the dataset with a large number of features, a large number of DNNs must be trained, leading to reducing the interpretability and increasing the computational resources required. In this work, we introduce a feature selection layer into NAM to reduce the number of DNNs in NAM to an arbitrary number and also to ease using two-input DNNs that can capture feature interactions. We verify the effectiveness of the proposed method in numerical experiments using benchmark datasets.