Research On Sequential Recommendation Algorithm Fusing User Information With Attention Mechanism

With the development of information technology and the Internet industry,resources are growing exponentially and information overload has become a major challenge.Personalised recommender systems can provide a promising solution to the above difficulties.In the field of sequence recommendation,sequence modelling as a graph structure has attracted extensive research.Aiming at the shortcomings of existing methods in capturing complex item conversion relationships and the fact that most studies do not clearly distinguish the impact of different historical clicks on current clicks,this paper proposes a new sequence recommendation algorithm(Graph Neural Network Fusing User Attributes Based on Attention Mechanism,A-UGNN).Firstly,a graph neural network sequence recommendation model that fuses user attributes(UGNN)is proposed to address the problem that previous GNN-based methods only focus on the information of each sequence itself,ignoring the personalised features of the users themselves.On the one hand,since different users have different behavioural patterns,which in turn lead to different item transformation relationships for each user,user information is injected into the GGNN module so that the user embedding and item hidden layer embedding can be fused to effectively extract personalised structural information from each subgraph at each node update.On the other hand,sequential graphing is missing location information,so the relative location information of each item is learned adaptively by merging the position embedding with the original item embedding,thus effectively exploiting the advantages of the sequential perceptron and graph perceptron models.Secondly,to compensate for the problem of user interest drifting over time,an attention mechanism is introduced based on UGNN,and a graph neural network sequence recommendation model based on an attention mechanism fusing user attributes(A-UGNN)is proposed.The last click in the original sequence is used as a query in the multi-head mechanism to adaptively obtain the current interest expression of the user,making it more sensitive to user interest drift than other GNN models.In addition,to make full use of the user and item feature information,static and dynamic preferences are fused to generate a unified representation of the user,where dynamic preferences represent the user's historical clicks and static preferences represent the user's inherent attributes.Finally,the proposed method is demonstrated to improve on the recommended accuracy metrics of Mean Reciprocal Rank(MRR),Normalised Discounted Cumulative Gain(NDCG)and Hits Ratio(HR)with two real data sets and outperforms other benchmarks.