A Benchmark Conservation Relevance Inference Method for Explaining Deep Networks with Multiple Feature Extraction Structures

Aiming at the problem that the counterfactual inference methods will cause relevance drift when interpreting the deep networks with multiple feature extraction structures, which leads to inaccurate interpretation, this paper proposes a benchmark conservation relevance inference method based on direct contribution (BCRI). By ensuring the consistency of the relevance propagation of multiple feature extraction structures, BCRI can overcome the relevance drift, and can accurately and quickly analyze the relevance of each input variable. This method can carry out the reasonable analysis of the model and understand the model behavior pattern. Experimental results show that the proposed method can generate more trustworthy counterfactual interpretations efficiently than other methods.