Research On Graph Neural Network Recommendation Algorithm Based On Multi-task Learning

With the development of Internet information technology,people need to process all kinds of information in their daily life,and information overload has become a major problem for people to quickly obtain the information they need.The recommender system has helped people greatly reduce the work of filtering information,whether it is for the work efficiency of business or the quality of daily life experience,bringing huge help and benefits.The implementation method of the recommender system is to mine the items that the user may be interested in based on the attribute information of the user or item,as well as the historical interactions between the users and items.However,traditional recommendation algorithms only focus on the direct interactions between users and items,but do not pay attention to the association of other attributes in the recommender system,and then have to face a lot of problems such as data sparsity.The modeling of the graph neural network on the heterogeneous information network can aggregate the information from other heterogeneous neighbor nodes to each target node through a specific relationship,and effectively mine the high-order implicit interaction between different types of nodes and edges.In addition,the practice of multi-task learning has proved its effectiveness for model improvement.Therefore,it is of great research value to apply graph neural network as a method to mine the interaction between users and items in recommender system under the framework of multi-task learning.This thesis proposes two multi-task graph neural network recommendation algorithms,which use graph neural networks to learn to generate node representation with rich semantics for recommendation.This thesis firstly proposes a graph neural network based on multi-task learning recommendation model,which was called Multi-Task Heterogeneous Graph Neural Network(MTHGNN).This model using the graph link prediction task and item matching recommendation task under the combined action of two supervised tasks learn users and items' node representation for recommendation.In order to further improve the model,this thesis proposes a graph neural network recommendation model,Contrastive Learning Heterogeneous Graph Neural Network(CLHGNN),which integrates contrastive learning on the basis of the MTHGNN model and let contrastive learning self-supervised task help the item matching recommendation task to learn node representations of users and items for recommendation.A self-supervised task-assisted item matching recommendation task using contrastive learning is studied to learn node representations of users and items for recommendation.In terms of algorithm design,the MTHGNN model first designs a dense subgraph sampling strategy for subgraph sampling of heterogeneous graphs;secondly,the graph neural network model includes the type conversion of heterogeneous nodes,the graph neural network module with attention mechanism and the mixed expert knowledge network module used in the parameter sharing of multi-task learning.Finally,in the multi-task joint training process,the automatic weight learning method is used to optimize the training process.The CLHGNN model adds additional content of contrastive learning on the basis of the above.First,in the self-supervised task module,the data obtained by subgraph sampling is enhanced to obtain new views samples from different perspectives;The self-supervised task module performs the new subgraphs coding,and the supervised task module also performs node representation;finally,before the recommendation prediction process of the supervised task,the node representation learned by the self-supervised task module and the node representation of the supervised task module are fully feature crossed for recommendation predictions.The experimental results of the two models on real public data verify their effectiveness in recommender system.