Detecting Statistical Interactions from Neural Network Weights

TL;DR

This paper introduces a new method for interpreting neural networks by directly analyzing their weights to detect statistical interactions, improving efficiency and accuracy over existing approaches.

Contribution

It presents a novel framework that identifies interactions from neural network weights without exhaustive search, leveraging the nonlinear activation effects and weight matrices.

Findings

Outperforms state-of-the-art interaction detection methods

Effective on both synthetic and real-world datasets

Highlights the importance of learned interactions in neural networks

Abstract

Interpreting neural networks is a crucial and challenging task in machine learning. In this paper, we develop a novel framework for detecting statistical interactions captured by a feedforward multilayer neural network by directly interpreting its learned weights. Depending on the desired interactions, our method can achieve significantly better or similar interaction detection performance compared to the state-of-the-art without searching an exponential solution space of possible interactions. We obtain this accuracy and efficiency by observing that interactions between input features are created by the non-additive effect of nonlinear activation functions, and that interacting paths are encoded in weight matrices. We demonstrate the performance of our method and the importance of discovered interactions via experimental results on both synthetic datasets and real-world application datasets.