Research On Heterogeneous Networks That Merge Meta-paths And Overlap Community Divisions Represent Recommendation Algorithms

With the rapid development of the Internet industry,a platform can always maintain the attractiveness to users and provide them with high-quality network service quality has become the primary task for major Internet manufacturers.Recommendation service is one of the most important tools to give money to the user experience and improve the stickiness of the platform to the user,because it can help users to select data information from a large number of data to meet the needs of users.The early recommendation system is based on the analysis of the scoring matrix,using the implicit feature representation of users and projects for recommendation.Because of its one-sided consideration,the data information obtained at the same time is often too sparse,resulting in the recommendation results are often unable to achieve the desired results.Therefore,many subsequent studies consider adding auxiliary information to the previous classical models to improve the recommendation results.In a real recommendation system,the interaction between users and projects(including purchase,collection and scoring)and the collection of attribute information between users and projects can be regarded as a network composed of different meanings of edges and nodes,that is,heterogeneous information network.Heterogeneous information network has attracted more and more attention in the field of recommendation because it can express complex information and capture more accurate feature representation in recommendation system.Most of the recommendation models based on heterogeneous networks first extract the eigenvector representation of nodes through the pre-set meta path,and then fit the scoring matrix to achieve scoring prediction.Although the existing methods of heterogeneous depression network have achieved some success in improving the accuracy,most of them still have the following problems:(1)Only using a single meta path to extract node information can not get accurate representation of nodes when faced with sparse data;(2)users' preference for different path semantics is ignored when fusing multiple meta path information,resulting in inaccurate representation of nodes in complex networks;(3)in dense data sets with large amount of interactive information,it is easy to introduce noise signals when fusing multiple meta path information(4)recommendation based on heterogeneous networks tends to take too much account of the presentation of biased structural association between nodes,while ignoring the tag similarity information containing specific semantics.1 ? In order to achieve the goal of introducing comprehensive structural information into the multi relationship network composed of massive data for accurate recommendation,aiming at the above four problems,this paper completed the following work:Aiming at problems(1)and(2),this paper proposes a recommendation algorithm for representation learning based on meta path network(ME-REC).We find that the users' preference for item scoring is strongly related to the hidden feature representation of Metapath.Therefore,we first use random walk strategy to obtain node sequences based on different Metapath,use metapath2 vec + + algorithm for different paths to learn thehidden vector representation between users and items in the unified dimension space,and then use MLP to learn the nodes for The preference weights of different meta paths are calculated,and the global representation vector is calculated.Finally,combined with the recommendation model of matrix factorization,the hidden factor vector decomposed from the scoring matrix is constrained by the hetesim similarity of the path structure,and the project scoring is predicted.Through the parameter analysis and comparison experiments in two widely used real datasets of different densities,movielens and Amazon,the parameter settings of the model when the performance of the model is optimal on different density datasets are obtained.In the comparison experiments,it is verified that the model has some improvement compared with the traditional matrix decomposition and the recommendation algorithm based on heterogeneous network,especially when the data is sparse Measuring accuracy.2?In order to further solve the problem that it is easy to introduce too much noise into the recommendation of merging multiple meta paths in data sets with high interaction density,that is to say,aiming at problems(3)and(4),this paper proposes a recommendation algorithm combining overlapping community partition and label similarity in multi relationship networks Similarity),or cpls-rec for short.Based on the concept of seed expansion and community label transfer in community division,we cluster users and project nodes in heterogeneous information networks,divide overlapping communities according to network structure information,and introduce label similarity as social regularization constraint,combined with classical matrix decomposition model for collaborative filtering and recommendation.Wefind that this method can effectively mine the community ownership of users according to their social relations and topological structure,which not only reduces the computational complexity of large-scale data sets,but also improves the accuracy of user preference extraction when the interaction information of user projects is sparse.Through the parameter analysis experiments on two real datasets with different densities,which are widely used in social network integration,doublemovies and yelp,the parameter settings of the model when the performance reaches the optimum under different proportion training sets are obtained,and then compared with other similar algorithms,it is proved that the model is more classical algorithm based on social network and recommendation calculation based on heterogeneous network Methods can have better performance,can improve the accuracy of project scoring,especially in the recommendation environment with high information density,can avoid the impact of noise data.