【摘要】 为解决当前基于会话的推荐系统方法在建模用户偏好时存在拍取的用户兴趣表示单一、静态问题,提出了一种动态层次意图学习网络。该网络同时考虑用户的多层意图和动态序列行为,设计了动态卷积神经网络和兴趣聚集门2个模块,并在每层抽取用户的特定粒度意图。此外还提出一个层级意图上下位损失函数,来约束用户意图的层次性。最后使用融合多种粒度的意图会话表示进行推荐。在3个真实数据集上的大量实验表明,模型在准确性和多样性上同时优于其他基于会话的推荐方法。更多还原

【Abstract】 To solve the problem of single and static representation of user interests when modeling user preferences in current session-based recommender system methods,a dynamic hierarchical intention learning network is proposed,which considers both the multi-layer intentions and dynamic user behaviors.Two modules,dynamic convolution neural network and interest cluster gate,are designed to extract users’ specific granularity intention in each layer.In addition,a constraint loss function is proposed to ensure the hierarchy of user intention.The final session representation incorporates multiple granularity intentions for recommendation.Extensive experiments on three real datasets show that the model outperforms other session-based recommendation methods in both accuracy and diversity.更多还原