Meta-Path Based Neighbors for Behavioral Target Generalization in Sequential Recommendation

Click-through rate (CTR) prediction is a crucial task in recommender systems, which aims to model users' dynamic preferences from their historical behaviors. To achieve this goal, most of the previous models adopt sequential neural networks (e.g., GRU) to encode the historical interactions into item representations for recommendations. Though these methods can perform well on recommending highly relevant items to users, we argue that such models are sub-optimal for the long-term user experience due to highly skewed recommendations: Monotonous items with similar subjects get more exposure because of inadequate interest explorations. Thus, some items which are not quite relevant to the users' historical preferences should be considered. To address these limitations, we propose a Heterogeneous Graph Enhanced Sequential Neural Network, HGESNN, to explore the interests of users beyond their historical interactions by explicitly modeling item relations with meta-path constructions. We incorporate a transformer-based network to embed personalized user intents into sequential learning. In the experiments on both public and industrial datasets, HGESNN significantly outperforms the state-of-the-art solutions. Specifically, HGESNN has been deployed in the main traffic of our Image-Text feed recommender system, which obtains 6.28%, 6.82%, and 4.77% CTR gains on news, novels, and entertainment contents, respectively.