MMNet: Multi-granularity Multi-mode Network for Item-Level Share Rate Prediction

Haomin Yu, Mingfei Liang, Ruobing Xie, Zhenlong Sun, Bo Zhang, Leyu Lin

Abstract: Item-level share rate prediction (ISRP) aims to predict the future share rates for each item according to the meta information and historical share rate sequences. It can help us to quickly select high-quality items that users are willing to share from millions of item candidates, which is widely used in real-world large-scale recommendation systems for efficiency. However, there are several technical challenges to be addressed for improving ISRP’s performance: (1) There is data uncertainty in items’ share rate sequences caused by insufficient item clicks, especially in the early stages of item release. These noisy or even incomplete share rate sequences strongly restrict the historical information modeling. (2) There are multiple modes in the share rate data, including normal mode, cold-start mode and noisy mode. It is challenging for models to jointly deal with all three modes especially with the cold-start and noisy scenarios. In this work, we propose a multi-granularity multi-mode network (MMNet) for item-level share rate prediction, which mainly consists of a fine-granularity module, a coarse-granularity module and a meta-info modeling module. Specifically, in the fine-granularity module, a multi-mode modeling strategy with dual disturbance blocks is designed to balance multi-mode data. In the coarse-granularity module, we generalize the historical information via item taxonomies to alleviate noises and uncertainty at the item level. In the meta-info modeling module, we utilize multiple attributes such as meta info, contexts and images to learn effective item representations as supplements. In experiments, we conduct both offline and online evaluations on a real-world recommendation system in WeChat Top Stories. The significant improvements confirm the effectiveness and robustness of MMNet. Currently, MMNet has been deployed on WeChat Top Stories.