Detecting Statistical Interactions from Neural Network Weights

A new preprint on detecting interactions from deep learning models:

Detecting Statistical Interactions from Neural Network Weights 

Michael Tsang Dehua Cheng Yan Liu 
University of Southern California 
May 16, 2017 

Abstract 

Interpreting deep neural networks can enable new applications for predictive modeling where both accuracy and interpretability are required. In this paper, we examine the underlying structure of a deep neural network to interpret the statistical interactions it captures. Our key observation is that any input features that interact with each other must follow strongly weighted connections to a common hidden unit before the final output. We propose a novel framework for detecting feature interactions of arbitrary order by interpreting neural network weights. Our framework, which we call Neural Interaction Detector (NID), accurately identifies meaningful interactions without an exhaustive search on an exponential solution space of interaction candidates. Empirical evaluation on both synthetic and real-world data showed the effectiveness of NID, which can uncover interactions omitted by other methods in orders of magnitude less time.