Research On Recommendation Algorithm Based On Graph Neural Network

In recent years,with the development of science and social progress,the Internet industry has developed rapidly.Although this change in the background of The Times has improved people's quality of life and accelerated the pace of social progress,it has also brought some troubles to people.In this era of information explosion,information overload is so severe that people cannot quickly and accurately find the part they want in the large amount of information.In order to solve the above problems,the recommendation system was born.The traditional recommendation algorithms are mainly divided into three categories: content-based recommendation algorithms,collaborative filtering recommendation algorithms and mix-based recommendation algorithms.Although these three types of recommendation algorithms have their own advantages and disadvantages,most successful recommendation systems now adopt collaborative filtering recommendation algorithms.The essence of collaborative filtering recommendation algorithm is to find user groups that have similar preferences with target users and recommend the items that the user groups are interested in to the target users.With the continuous improvement of the recommendation algorithm,people combined the traditional recommendation algorithm with the deep learning method in the later stage and put forward many recommendation algorithms with better performance.Collaborative filtering recommendation algorithm,as the most popular recommendation algorithm today,has shown excellent performance in various fields,but there are still some shortcomings.As the number of users and projects continues to increase,the problem of data sparsity arises.One of the most successful collaborative filtering recommendation algorithms is based on matrix decomposition.However,the potential embedding representation of matrix factorization learning is transgenic in nature and is not intended to be generalized to potential users and projects.To solve the above problems,this paper proposes a graph neural network model recommendation algorithm based on matrix completion and noise reduction.Firstly,aiming at the sparse problem of the original scoring matrix data,Singular Value Threshold(SVT)algorithm is used to prefill the sparse scoring matrix to alleviate the sparse problem of the scoring matrix data.Then,based on the prefilling of the scoring matrix data,Singular Value Decomposition(SVD)is used to de-noise the prefilling data.Finally,on the basis of the denoising matrix,an inductive graph neural network refinement model is introduced to make recommendations.Aiming at the disadvantage that the inductive graph neural network model can't realize the Adjacent Rating is closer to the fact,the algorithm carries out Adjacent Rating Regularization(ARR)constraint in the loss function.In order to verify the effectiveness and feasibility of the proposed algorithm,a comparative experiment is conducted on the Movielens-L00 K dataset and Douban score dataset.The experimental results show that the proposed algorithm has excellent recommendation ability and migration ability.In conclusion,SVT and SVD were used in matrix completion and noise reduction,and ARR was introduced into the inductive graph neural network model to solve the problem that it could not distinguish the similarity between different ratings.The excellent recommendation ability and migration ability of the proposed algorithm were verified through experiments.