DeepROCK: Error-controlled interaction detection in deep neural networks

TL;DR

DeepROCK is a novel method that uses knockoffs and a specialized DNN architecture to reliably identify feature interactions with controlled false discovery rate, aiding interpretability in complex neural networks.

Contribution

DeepROCK introduces a systematic approach for interaction detection in DNNs that controls FDR and enhances interpretability, addressing limitations of existing methods.

Findings

Effectively controls false discovery rate in interaction detection

Demonstrates superior performance on simulated datasets

Validates approach on real-world data

Abstract

The complexity of deep neural networks (DNNs) makes them powerful but also makes them challenging to interpret, hindering their applicability in error-intolerant domains. Existing methods attempt to reason about the internal mechanism of DNNs by identifying feature interactions that influence prediction outcomes. However, such methods typically lack a systematic strategy to prioritize interactions while controlling confidence levels, making them difficult to apply in practice for scientific discovery and hypothesis validation. In this paper, we introduce a method, called DeepROCK, to address this limitation by using knockoffs, which are dummy variables that are designed to mimic the dependence structure of a given set of features while being conditionally independent of the response. Together with a novel DNN architecture involving a pairwise-coupling layer, DeepROCK jointly controls the false discovery rate (FDR) and maximizes statistical power. In addition, we identify a challenge in correctly controlling FDR using off-the-shelf feature interaction importance measures. DeepROCK overcomes this challenge by proposing a calibration procedure applied to existing interaction importance measures to make the FDR under control at a target level. Finally, we validate the effectiveness of DeepROCK through extensive experiments on simulated and real datasets.