融合实体与关系交互信息的知识感知推荐模型

Knowledge-aware recommendation model fused with interaction information between entities and relations

由于知识图谱包含了丰富的项目属性及其关联信息,因此在推荐系统中引入知识图谱能在一定程度上解决数据稀疏和冷启动问题。如基于传播的推荐系统就利用了知识图谱的图结构学习用户及项目表示等相关特征。但在传播过程中,往往忽略了实体与关系之间的交互信息对特征表示的贡献,由此提出一种融合实体与关系交互信息的知识感知推荐模型。首先,将协同信息和知识关联整合,采用异构传播方式传播并扩展用户和项目的表示。其次,在传播过程中用注意力机制强化实体与关系之间的交互信息,增强语义关联,保证用户和项目基于知识的高阶交互的有效性。然后采用知识感知注意力机制来区分每层中实体邻居的重要性,更精确地生成用户和项目的表示。最后通过聚合器将多个表示结合得到用户和项目的最终表示,从而预测用户与项目进行交互的概率。通过添加KL散度损失函数对模型进行优化,以对齐模型的预测分布和真实分布之间的差异。在Last.FM、Book-Crossing和MovieLens-20M 3个数据集上进行的实验结果表明该模型在CTR预测性能中比其他基线模型有较大提升。

As knowledge graphs contain rich item attributes and their associated information,introducing knowledge graphs into recommendation systems can to some extent solve data sparseness and cold start problems.For example,recommendation systems based on propagation utilize the graph structure of knowledge graphs to learn relevant features such as user and item representations.However,the contribution of interactive information between entities and relationships to feature representation is often ignored in propagation,so this paper proposed a knowledge aware recommendation model that fused with interaction information between entities and relationships.Firstly,collaborative information and knowledge correlation were integrated,and heterogeneous propagation methods were used to propagate and expand the representation of users and items.Secondly,in the process of propagation,attention mechanism was used to strengthen the interaction information between entities and relationships,enhance semantic relevance,and ensure the effectiveness of knowledge-based high-level interaction between users and items.Then,a knowledge aware attention mechanism was used to distinguish the importance of entity’s neighbors in each layer,and generate representations of users and items more accurately.Finally,to predict the probability of user interaction with item,multiple representations were combined to obtain the final representation of user and item by an aggregator.To optimize the model,KL divergence loss function was added to align the difference between the prediction distribution and the real distribution of the model.Experimental results on three datasets of Last.FM,Book-Crossing and MovieLens-20M show that the proposed model has a great improvement in CTR prediction performance compared with other baseline models.