Quantifying Explanations of Neural Networks in E-Commerce Based on LRP

Neural networks are a popular tool in e-commerce, in particular for product recommendations. To build reliable recommender systems, it is crucial to understand how exactly recommendations come about. Unfortunately, neural networks work as black boxes that do not provide explanations of how the recommendations are made. In this paper, we present TransPer, an explanation framework for neural networks. It uses novel, explanation measures based on Layer-Wise Relevance Propagation and can handle heterogeneous data and complex neural network architectures, such as combinations of multiple neural networks into one larger architecture. We apply and evaluate our framework on two real-world online shops. We show that the explanations provided by TransPer help (i) understand prediction quality, (ii) find new ideas on how to improve the neural network, (iii) help the online shops understand their customers, and (iv) meet legal requirements such as the ones mandated by GDPR.