Recently, the recommender system combining knowledge graph (KG) is gaining traction from researchers. However, current research faces two problems: The first is most existing works ignore the fact that real-world knowledge graphs are often noisy and contain lots of recommendation-irrelevant connections. The second is item-linked relations can reflect different attributes of items, however, previous works usually ignore the different relational semantics carried by different types of relational paths, leading to suboptimal recommendation performance. To address these issues, we propose a novel model: Sampling on Knowledge Graph for Recommendation with Relational Path-aware Graph Neural Network (SKRAG). We design a relational path-aware sampling method to extract recommendation-relevant information in KG, further mitigating the impact of KG noise; Then, we utilize relational path information to capture item long-range knowledge associations with explicit attribute semantics and extract user potential purposes for learning high-quality item and user representations. We conduct extensive experiments on three datasets, and the experimental results show the superiority of our model.

The session-based recommendation(SBR) aims to predict user actions based on anonymous sessions. SBR focuses on recent sessions to make more real-time and reliable recommendations and does not need to extract user IDs, thereby alleviating privacy issues. Recent research mainly models user preferences based on the target session while ignoring other sessions, which may contain items transitions related to and unrelated to the target session. This paper proposes a novel method, namely Improved Global and Local Graph Neural Network (IGL-GNN), to models item transitions within not only the target session but also the other sessions. Specifically, we learn the session-level item representation from the local graph, learns the global-level item representation from the global graph. We propose a new method to effectively aggregate global and local representations, which not only retains more useful information but also removes useless information. In addition, we innovatively introduced Transformer, incorporated it into the model as a general deformation function, enhanced the ability to obtain complex transformations, and solved the limitations of the model in learning complex representations. We have done a lot of experiments on three benchmark datasets, and the results show that the performance of IGL-GNN is better than the most advanced methods.

In recent years, more and more recommendation algorithm considers social information. However, in the existing social recommendation algorithm focused on the rating prediction task, the user's rating bias and the item's rating bias are often not considered. Besides, the existing social recommendation algorithms usually fail to deal with the sparse problem of social information. Concerning the problem of ignores rating bias and social information sparsity in the social recommendation, this paper proposes a novel recommendation model that combines implicit information and rating bias based on the traditional matrix factorization model. The social information used includes explicit social information and implicit social information generated by collaborative user network embedding methods. Experimental results on two real public datasets show that compared with other recommendation models, the proposed model reduces 1.13% to 10.09% and 3.14% to 8.38% respectively in the Root Mean Square Error (RMSE) and the Mean Absolute Error (MAE). It shows that our proposed recommendation algorithm has a better recommendation effect.