Self-Supervised Hypergraph Attention Networks For Session-Based Recommendation

Recommendation systems process the interaction data between users and items to make personalized recommendations to users.Session-based recommendation is an important branch of recommendation systems,which uses the user’s behavior in a short period of time to predict the next most likely item to be interacted with,and is real-time and dynamic.Through the analysis of existing recommendation methods,it is found that there are still three problems in the current session recommendation methods: First,in a short session,it is not easy to make efficient recommendations using only the information of the session sequence itself,and the potential information of the graph nodes constructed by the session cannot be fully utilized,ignoring the complex higher-order information among items.Second,the process of obtaining session information does not consider the degree of influence of click item information on the recommendation results;third,during the training process of the deep network,the model is prone to the problem of oversmoothing,and the ability to extract information from higher order is limited,and the long-range session information is also missed in the propagation process.To address the above problems,this thesis proposes a self-supervised session recommendation model based on hypergraph attention networks to enhance session recommendation.The main work is as follows:Firstly,the complex higher-order relationships among items are captured by modeling session-based data as hypergraphs.Modeling each session data as a hypergraph,different hypergraphs are connected by shared entries to form a hypergraph containing item-level higher-order associations.In addition,the inherent data sparsity problem may also limit the benefits from hypergraph modeling.To address this issue,a line graph describing session-level relationships is built on the basis of the hypergraph by modeling each hyperedge as a node with a focus on the connectivity of the hyperedges.After modeling the same session data into two different views,an item-level hypergraph describing intra-session information and a session-level line graph describing inter-session information,it is used for downstream task network model propagation.Secondly,a self-supervised session recommendation model based on hypergraph attention networks is proposed.First,the model uses hypergraph attention network layer channels to encode and propagate the item-level session hypergraph data that has been modeled,and input the session hypergraph data into the item-level hypergraph convolution layer to obtain a vector representation of session click items that contains more potential information.The vectors are then fed into the hypergraph attention layer to learn the dynamic correlation matrix and assign corresponding weights to different click item vectors,ignoring the influence of irrelevant information on the model recommendations.At the same time,the line graph constructed on the basis of the session hypergraph is fed into the line graph convolutional network channel to learn to the inter-session embedding representations.In this way,the two channels in the network can be viewed as two different views describing intra-session and inter-session information,but at this point the two channels are relatively independent and know little information about each other.Therefore,self-supervised learning is incorporated into the network training to maximize the mutual information between the learned session representations of the two channels to obtain the final recommendation results.Thirdly,to address the problem that the model is prone to oversmoothing when the layers become deeper and the long-range historical session information is also missed in the propagation process,we propose to add residual connections to the model to enhance the generalization ability of the model,and to ensure the updating and learning of network parameters in deeper networks through residual connections.To verify the effectiveness of the model proposed in this thesis,the model is compared with other baseline models accordingly,and the experimental results show that the model achieves notable results on three benchmark datasets: Tmall,Nowplaying,and Diginetica.