Relevance aggregation for neural networks interpretability and knowledge discovery on tabular data

Abstract

The lack of interpretability of neural networks is partially why they are not adopted in a wider variety of applications. Many works focus on explaining their predictions, but few take tabular data into consideration, which led to a small adoption even though this data is of high academic and business interest. We present relevance aggregation, an algorithm that combines the relevance computed from several samples as learned by a neural network and generates scores for each input feature. We also present two methods for visualizing the learned patterns, leading to a better model comprehension. The method was tested in synthetic and real-world datasets (breast cancer gene expression, online shopping behavior, and national high school exam) for classification and regression tasks. It correctly identified which features are the most important for the network’s predictions. The selected features can be distinct for each class. The rank of features scores also matches their contribution to the model’s performance. The results selected relevant features from the data, paving the way for knowledge discovery. The top-ranked features were consistently able to improve the performance of another independent classifier. For poorly trained neural networks, relevance aggregation helped identify incorrect rules or machine bias.