Research On Matrix Decomposition Recommendation Algorithm Based On GNN

The previous collaborative filtering recommendation algorithm is to model the historical interaction behavior of users and commodities to mine user interest preferences,classify their preferences,and match similar items(or items)they are interested in.However,with the continuous expansion of the number of users and items on the network,the resulting data sparseness and low recommendation accuracy have gradually become important factors that limit the research progress of collaborative filtering recommendation algorithms.In view of this,based on the relevant research results and based on the personal knowledge and skills,this paper plans to construct a collaborative filtering recommendation algorithm composed of graph neural network and matrix decomposition.Combine the flexibility and extensibility of the matrix factorization model with the performance advantages of graph neural networks on heterogeneous graphs to improve the accuracy of the model.In the application process of this algorithm,the graph neural network is first used to model the interaction graph between users and projects,and the social network graph between users and users,and train users' potential characteristics in the two dimensions of user space and project space.Fully capture the user's interest preferences;secondly,introduce a two-layer attention network training weight parameter to increase model flexibility;finally integrate the learned user potential feature model into the matrix decomposition as a constraint item on the user side to reconstruct the scoring matrix,Which produces the final prediction result.The comprehensive experimental analysis on two public datasets ciao and epinions shows that the root mean square error and average absolute error of the proposed model are reduced to varying degrees compared with the original PMF,and the reduction rates are 2.91%,3.10%and 4.83%,3.84%respectively.Compared with other advanced models,the proposed model also shows superior recommendation performance.This paper has 29 pictures,11 tables,and 67 references.