Research On Information Recommendation System Based On Community Discovery In Social Network

With the rapid development of the scale of the Internet,the amount of information disseminated on the Internet has increased rapidly,and how to perform preferential information retrieval in large-scale information has become an important research topic.In order to meet the needs of users for information push,recommendation algorithms came into being.Early recommendation algorithms can generate an interaction matrix based on the user’s interaction history with items,and calculate recommended content that meets user preferences based on the interaction matrix.This method is still used by most recommendation models to this day.However,the interaction matrix between users and items in the real environment is extremely sparse,and the interaction items in the matrix generally only occupy 1%to 4% of all potential interactions,and as the number of users and items continues to increase,this sparsity will also become It’s getting worse.For recommendation algorithms based on interaction matrix to mine information,the sparseness of the matrix severely limits the performance of the recommendation system.Since the algorithm relies on the existence of the interaction matrix,the cold start of the system will make the model difficult to train.In addition,the recommendation algorithm based on the interaction matrix has a poor ability to fit the data due to its own structural limitations,and the performance bottleneck is obvious.Therefore,how to solve the matrix sparsity and cold start problems and enhance the model’s ability to fit data has become an important direction to improve the performance of recommendation algorithms.With the development of machine learning models in recent years,linking machine learning models with recommendation systems provides the possibility to solve the above problems.Considering the sparse characteristics of the interaction matrix,this paper considers adding knowledge graphs as a supplement to training information,and proposes a joint learning network model SNATMF that integrates attention community aggregation and matrix decomposition technology,and proves its performance in a large number of experiments and analysis verifications.The advantages.The main work of this paper is as follows:(1)Transform the knowledge graph into a topological network structure,and propose an algorithm for community neighborhood aggregation of the network structure in the graph neural network.This algorithm can aggregate information in the high-dimensional neighborhood of the target node to obtain an embedding vector containing rich semantic information.Such embedding vectors can provide sufficient information supplement for recommendation to alleviate the effect of interaction matrix sparsity.(2)Inspired by the attention mechanism,a dual-ended attention mechanism is incorporated into the above aggregation process,and the biased aggregation of the target node in the community neighborhood is performed on the user end and the project end respectively.The addition of the attention mechanism enables the aggregation process to consider user preferences to obtain more comprehensive weight settings and improve the embedding effect of the aggregation.(3)Generate latent vector representations of users and items by decomposing the interaction matrix,and use it as a supplement to the embedding vector feature space.This method can predict the user’s potential interaction score for unknown items,which can improve the utilization rate of the model for the interaction matrix,and further alleviate the impact of the sparse interaction matrix on the performance of the model.This paper conducts detailed experiments on the model on the Movielens-1m dataset.By comparing the experimental data with other six benchmark algorithms,the performance advantages of the model in this paper are proved in terms of CTR prediction and Top-K prediction.At the same time,for the model itself,this paper discusses the influence of different hyperparameters on the model through control experiments on hyperparameters,and obtains the optimal combination of hyperparameters.Finally,this paper conducts ablation experiments on different modules in the model,and analyzes the influence and importance of different modules on the performance of the model.Compared with the recommendation algorithm based on interaction matrix,SNATMF has stronger fitting ability and can better solve the problems caused by matrix sparsity,and has better predictive performance.For other recommendation algorithms that incorporate knowledge graph information,SNATMF also has a good performance improvement.