基于双图混合随机游走的社会化推荐模型

Bi-Graph Mix-random Walk Based Social Recommendation Model

近年来,可以有效缓解数据稀疏和冷启动问题的社会化推荐受到了研究者和业界的关注.社会化推荐利用显式或隐式社交关系作为辅助信息,提升了推荐性能.然而,目前的社会化推荐模型通常采用普通图描述社交关系.普通图中的边常描述为成对节点的关系,这种方法适合描述显式关系,但难以描述复杂的隐式关系,如购买过同一商品的多个用户之间的集合关系,因此难以学习到准确的节点表示,影响推荐的性能.针对此问题,本文结合超图和普通图,提出基于双图混合随机游走的推荐(BG-Rec)模型.构建超图描述复杂的隐式关系,同时用普通图描述显式的社交关系,并在两种图上定义混合随机游走策略,生成结合隐式关系和显式关系的游走节点序列,学习更准确的节点嵌入表示.根据用户评分的高低,构建了正反馈超图和负反馈超图,考虑更细粒度的朋友关系,以识别可靠的朋友.融合可靠朋友的偏好和后验概率最大化优化物品个性化排序.三个公开数据集的大量实验表明了BG-Rec在推荐性能上的优越性,冷启动和消融实验表明了其在缓解冷启动问题的有效性和超图建模的合理性.

In recent years,social recommendation approaches have attracted attention because they can effectively improve the recommendation quality when user-item interaction data is sparse.Explicit and implicit social relations,as auxiliary information,are used to improve the recommendation quality.However,social relations are represented by simple graphs in existing models.The nature of edges connecting pair-wise nodes in simple graphs makes it suitable for describing explicit relations.Still,it is incapable of modeling complex implicit relations,such as the collective relation between multiple users who have purchased the same product.Therefore,it isn't easy to learn the node representation accurately,only based on simple graphs,which even affects the recommender's performance.In this paper,we propose a recommendation model based on a bi-graph hybrid random walk(BG-Rec)to overcome this problem,which combines hypergraph and graph.We construct a hypergraph and a simple graph to depict complex implicit relations and explicit social relations separately.Next,the mixed random walk strategy(MixRandom)is used to generate node sequences that combine implicit and explicit relations.Furthermore,node sequences are used for learning more accurate representations of nodes.Then,positive feedback hypergraph and negative feedback hypergraph are constructed based on user ratings,so that more fine-grained friend relations can be considered to identify reliable friends.Finally,the personalized ranking of items is optimized by considering the preferences of reliable friends and the maximization of the posterior probability.Experiments on three public datasets show the superiority of BG-Rec in recommendation performance.The cold-start study and ablation study validates the effectiveness of alleviating the cold-start problem and rationality of hypergraph modeling.