细粒度建模用户兴趣的序列化推荐方法

Fine-grained modeling of user interests for sequential recommendation

序列化推荐因其实用性和较高推荐精度在近期受到了人们广泛关注.不同于传统推荐方法,序列化推荐的核心在于如何基于用户近期交互行为来捕获用户的短期兴趣.现有工作或者依次考虑用户交互序列中物品之间的成对关系,忽略了更为重要的多对一关系;或者将用户最近交互的多个物品视为集合,忽略了物品之间的相对次序.这不仅未能充分挖掘物品之间的复杂关系,而且未准确刻画用户兴趣的演变过程.为此,本文提出门耦合胶囊网络(gating coupled capsule network,GCC),一种从个体层、联合层以及局部有序性3个层面细粒度分析物品间关系对于用户短期兴趣影响的方法.借助胶囊网络的空间感知能力,GCC引入个性化胶囊模块建模用户高阶时序信息,其不仅能够捕获物品间的联合层关系,也保留了物品间的局部有序性.另外,本文在GCC中设计了用于建模物品间成对关系的个性化门单元模块,以此捕获个体层关系对于用户短期兴趣的影响.在4个真实推荐场景下的实验结果表明,GCC相比于主流序列化推荐方法在多个评价指标上具有显著的性能优势.

In recent years, the sequential recommendation method has attracted extensive attention owing to its practicability and high accuracy. Different from general recommendations, explicitly capturing short-term interests based on users’ recent actions lies at the heart of a sequential recommender. Existing methods have either modeled successive items step by step without considering the many-to-one relationship or neglected the local-order information, considering the recent items as a set. This neither fully mines complex item transitions nor depicts the evolution of user interests. Therefore, we propose the gating coupled capsule network(GCC),a fine-grained analysis method to model individual-level, union-level, and local-order relationships of short-term user interests. Specifically, we designed a user-specific capsule module to capture high-level temporal interaction sequences, which model union-level patterns and perceive local-order relations. Moreover, we present a personalized gating module to focus on the pairwise relationships among items to capture the influence of individual-level information on short-term user interests. Extensive experiments under four real-world recommendation scenarios demonstrate that GCC outperforms the state-of-the-art methods on different ranking metrics.