How to Find It Better?: Cross-Learning for WeChat Mini Programs

He Li, Zhiqiang Liu, Sheng Xu, Zhiyuan Lin, Xiangqun Chen

Abstract: WeChat Mini Program is a lightweight app relying on the WeChat client, which can be accessed directly from the search list without downloading and installing. Retrieval and ranking for the Mini Programs differ from traditional web search in two sides. On the one hand, as the search queries are often short and most Mini Programs contain few useful textual information, it is hard to retrieve when the user input is inaccurate. On the other hand, without user scoring and rating system like App Store and Google Play, it is hard to rank relatively better results in a more advanced position. In this paper, we propose a Cross-Learning strategy to improve the search experience, where the semantics of queries and Mini Programs are represented not by itself, but by each other. We treat the search task as an extreme multi-label classification problem where the queries are inputs and the Mini Programs are labels. We propose a N-Gram self-attention query encoder to capture the search intention behind these short queries, and carefully design the label selection strategy based on user behavior to rank higher quality Mini Programs in higher positions. Our model outperforms some state-of-the-art baselines in the offline environment, and brought improvement to our actual business in the online A/B Test, which proves the practical significance of our work.