Research On Sequential Recommendation Based On Graph Neural Networks

Sequence recommendation is an important branch of recommender system research.Most of the existing methods can't effectively utilize item relations and attribute infor-mation in the data.The influence of the time interval and co-occurrence frequency between items on the item relations is ignored.In addition,the existing methods ignore the influence of common attributes between items,and can't distinguish between item attributes.To solve the above problems,this paper proposes sequential recommendation models based on graph neural networks to effectively extract the above information.Firstly,a sequence recommendation model that fuses relational information is pro-posed.The model consists of a relational information extraction module and a relational information fusion module.The time interval and co-occurrence frequency between items are used as edge weights of the item graph.The edge weights reflect the strength of the item relations.In the user graph,edges are added between each user and its' similar users.Then,the relational information extraction module obtains relational information from the user graph and the item graph through the graph convolutional network.The relational information fusion module uses the obtained relational information to enhance the representation of item sequences and model the long-term interest of users.Empiri-cal results on five datasets show that the proposed model significantly outperforms other baselines,in average up to 8.33% and 6.15% absolute improvements in Recall@10 and NDCG@10,respectively.Based on the first work,a sequential recommendation model that fuses attribute information is proposed.The model utilizes the attribute information of items in two ways.On the one hand,edges are added to the item graph for pairs of items with common attributes.The attribute-level item relations can enrich the relational information of items.On the other hand,the graph attention network is used to distinguish between item attributes and learns item representation in the attribute space to enhance the features of the sequence.Empirical results on two datasets which include item attributes show that the proposed model significantly outperforms other baselines,in average up to 20.88%and 13.74% absolute improvements in Recall@10 and NDCG@10,respectively.To alleviate the impact of data sparsity on the performance of sequence recom-mendation,this paper combines contrastive self-supervised learning with two proposed models and proposes a contrastive self-supervised sequence recommendation frame-work.This framework further strengthens the representation of relational information and attribute information by establishing contrastive learning tasks for the relational information and attribute information obtained in the graph.Compared with the two original models and other contrastive self-supervised sequence recommendation meth-ods,two models adopting this framework achieve an average improvement of 9.04%and 12.79% on NDCG@10,respectively.