Towards Explainable Search Results in E-commerce

Search result explanations are essential in E-commerce, helping users understand the relevance of the returned results. Existing methods primarily focus on explaining relevance based on either product content or behavioral data. However, we argue that combining both content and behavior data can provide more comprehensive and accurate explanations. In this paper, we propose a novel approach to generate relevance explanations. First, we utilize the content data to train a domain-specific large language model (LLM) that generates relevance labels and reasoning processes for queries and items. Then, we introduce the BehaviorRAG framework to retrieve behavioral data related to queries and items, allowing the model to generate explainable reasons for their relevance. Finally, the LLM integrates outputs from both the content- and behavior-based modules to produce a final explanation. To evaluate the effectiveness of our methods, we conduct extensive experiments on both our built dataset and publicly available datasets. The results demonstrate that our method outperforms current state-of-the-art baselines in predicting relevance and generating explainable reasons. Furthermore, online A/B testing on the Fliggy app demonstrates that incorporating explanations generated by our approach into the search results leads to a 2.30% increase in the Unique Visitors List to Order. The codes are publicly available at https://github.com/tonnyaudio/explainable.