基于影响力扩散的会话推荐模型

会话推荐的任务是根据用户近期的点击行为预测下一个点击.该领域之前的模型主要关注到会话推荐中的时序模式(序列特征),但是由于用户兴趣迁移以及商品多属性等因素,物品之间的交互关系可能比呈现出来的时序模式更为复杂.为了解决该问题,受PageRank算法的启发,把会话点击和网页跳转联系起来,提出了一个会话推荐中的影响力扩散模型.具体地说,该模型在会话序列的显式时序结构之上构建了会话图,刻画出更加丰富的转移路径,并通过图扩散模型捕获到物品之间的潜在交互关系.在会话表示阶段,该模型提出了一种新颖的位置编码方式来应对兴趣迁移的状况,并在此基础上设计了一种意图提取框架,能在多兴趣会话中迭代出核心意图.在真实数据集上的实验结果表明,所提模型相较于以往方法有较好的性能,并有效解决了兴趣迁移的问题.

Session recommendation model based on influence diffusion

Session-based recommendation aims to predict the next click based on the user''s current click behavior. Previous methods mainly focus on the time-series pattern(i. e. sequential pattern) in session, nearly ignoring the complex transition among items. To this end, inspired by PageRank, this paper proposed an influence diffusion model, which built the bridge between session clicks and web page jumps. Specifically, this method built a session graph based on the explicit time-series pattern of the session sequence, which could model more meaningful transition patterns. Furthermore, it applied a graph diffusion model to capture the potential interactions among items. In the process of session representation, this method proposed a novel position encoding method to tackle the problem of interest shift and an intention extraction layer on top of it, which could extract core intention in multi-interest session. The experimental results on real-world data demonstrates the advantages of this method over state-of-the-art methods.