Meta-path And Graph Structure Augmented Graph Neural Network For Social Recommendation Research

With the booming growth of online social information,easily accessible social information has been widely used in recommender systems to alleviate the data sparsity problem in traditional collaborative filtering.However,the question of how to adequately integrate social information into the model so that the final embedding representation can be improved has been the subject of much research.As users play an important role in both social networks as well as user-item interaction networks,many current studies have modelled the process of information diffusion on both networks through graphical convolutional neural networks,followed by information fusion to obtain higher-order information.However,three problems with this modelling approach are that firstly,some potential collaboration signals are not explicitly encoded,a practice that may lead to some higher-order collaboration signals not being captured,secondly,some unreliable interactions may also introduce noise leading to performance degradation due to the unreliability of links,and thirdly,there may be some unobserved positive interactions in a fixed graph structure due to data sparsity.The following work has been done in this paper to address the above issues.(1)To address the problems of inability to capture higher-order collaboration signals and unreliable links,this paper proposes Meta-path Enhanced Lightweight Graph Neural Network for Social Recommendation(ME-LGNN).The social network and the user-item interaction network are first fused into a complete heterogeneous information network on which higher-order collaborative signals are captured through explicit coding by a lightweight graph neural network.Then,in order to enable users to capture reliable information more efficiently,a series of interpretable meta-paths are designed to constrain the model and further enhance the embedding representation by constraining the dependency probabilities of the meta-paths to make the more closely related nodes more closely connected to each other.(2)In order to be able to mine positive preferences in unknown preferences,this paper propose Graph Enhanced Lightweight Graph Neural Network for Social Recommendation(GE-LGNN),which designs an enhanced graph module.The graph structure and representation learning ability are enhanced iteratively through the graph update module and the representation learning module,the sparsity of the data is alleviated and better recommendation performance is obtained.(3)In this paper,extensive comparison experiments are conducted with several existing classical recommendation models on three common datasets.The experimental results validate that the two social models proposed in this paper are effective in improving the recommendation performance.It also confirms the effectiveness of the proposed social recommendation model in dealing with the data sparse problem.