Research On Social Recommendation Algorithm Based On Graph Neural Network

Aiming at the problems existing in the recommendation algorithm,this paper studies recommendation methods in the social recommendation scenario,and designs two graph neural network-based social recommendation algorithm frameworks to achieve the purpose of improving the recall rate and optimizing the sorting results.The main research contents of this paper are as follows:1.Graph multi-attention network-based for social recommendation.Graph neural network technology has a wide range of applications in the field of social recommendation.However,as the diffusion depth increases,it tends to lead to over-smoothing issues that inhibit its performance.This paper proposes a similarity-based multi-relational attention network for social recommendation scenarios.The proposed model has three salient features:(1)it tries to alleviate the data sparsity problem in socialized recommendation scenarios by using social relations among users and homogeneous relations among items as supplementary information;(2)It has an iterative aggregation structure.In user and item modeling,four aggregation operations are introduced to deal with two different diffusion processes respectively,capable of imitating higher-order interest diffusion in the user-item domain,and higher-order influence diffusion in the user and item domains;(3)It designs two attention mechanisms capable of distinguishing importance weights when building user and item embeddings.Specifically,node-level weights indicate the strength of each interactive connection,and graph-level weights focus on how to balance social influence and interest influence.Experiments on two representative large-scale datasets show that the algorithm significantly outperforms previously proposed methods,with gains of 5.7% HR@10 and 6.8% NDCG@10,respectively,over the optimal baseline.Furthermore,the proposed model can alleviate the over-smoothing problem,and its performance can be further improved by increasing the diffusion depth.2.Gated graph sequence neural network-based for social recommendation.The session-based recommendation problem aims to predict the next behavior of users within a session,and previous methods only exploit the transition information of adjacent items in the session sequence to build a sequence model,which is not enough to capture the frequent co-occurrence behaviors in a session.To obtain more powful session embeddings and consider complex transition patterns between items,this paper proposes a session-based graph attention network for social recommendation scenarios.The proposed model has three salient features:(1)It models the session sequence as a session graph,and uses the gated graph sequence neural network to model the target user’s interactive behavior sequence in the current session,extracts the local and global information of the session graph,and obtains the user’s dynamic preference representation.(2)It distinguishes short-term preference and long-term preference in social influence,each friend’s short-term preference is represented by the graph embedding of its own session graph,and the long-term preference is encoded into a learnable personal embedding vector.Integrate the above-mentioned long-term and short-term representations using nonlinear transformation to obtain the friend’s final preference representation.(3)It combines the preference representation of the current user with that of its friends,and uses an attention mechanism to distinguish the social influence contributions of different friends,and obtain contextdependent social influence information.The method can be effectively extended to large datasets,and the experimental results on three public datasets show that the accuracy of the model is significantly better than that of all state-of-the-art baselines,exhibiting good recommendation performance.