Personalized Recommendation Using Matrix Decomposition And Implicit Relations

The collaborative filtering algorithm of recommendation system has been widely concerned and made great progress,and the matrix decomposition method plays an important role in collaborative filtering technology.In order to establish rec-ommendation information,collaborative filtering technology needs to connect users and items,which are essentially different entities.The key to the con-nection lies in the historical behavior data of users,which is usually expressed as the relationship matrix between users and items.The performance of recom-mendation system depends on the input data.The most efficient data is the user's explicit feedback,that is,the direct expression of interest.The traditional matrix decomposition method is mainly to decompose the original matrix into two matrices of users and objects,map the two entities with essential differences into the same cryptic space,and explain the links and scores by describing the characteristics of the two entities in the cryptic space.Because explicit informa-tion is not often available,an important direction of improving the model is to use rich implicit feedback to indirectly infer users,preferences.There are many types of implicit feedback.In recent years,there have been many researches on how to measure the importance of these information in rec-ommendation,and they have been successfully applied to the model.To produce more accurate recommendation results,we need to identify all available features in the data.Using more available features can solve the problem of sparsity of relational matrix,and improve the scope of application of the model.However,implicit feedback can not directly reflect the user's preferences,so in practical application,it is necessary to give priority to explicit feedback information,sup-plemented by implicit feedback information,and comprehensively use various types of data to capture the complex interaction between users and objects.Ma-trix decomposition is widely used because of its excellent scalability,which can find a balance between explicit feedback and implicit feedback.This paper shows the new application of matrix decomposition technology in personalized recommendation,and further expands the applicable scenarios of matrix decomposition and improves the accuracy of recommendation by inte-grated implicit relations.We found that there are other entities that affect the recommendation results besides the two entities of users and goods.Comprehen-sive utilization of these hidden relationships between multiple entities can improve the accuracy of recommendation.By introducing the implicit relationship into the model,we find out new features.In the supply chain of the enterprise net-work,we mine the higher-order connection structure as the feature,and propose the enterprise partner recommendation algorithm embedded in the higher-order connection structure.For the movie scoring prediction scenario,we mine the user's interest preference information of the movie crew members,and propose the integrated implicit relationship movie recommendation algorithm.Our re-search finds out the implicit feedback information through further analysis of the problem,makes full use of the complex relationship involving multiple entities and mines new features,which not only alleviates the highly sparse problem of the original relationship matrix,but also improves the accuracy of recommenda-tion,and expands the use scenario of the recommendation model using implicit feedback.The main contributions of this paper are as follows:1.We find that in the complex network with multiple roles,the higher-order connection structure reveals the implicit relationship in the network evolu-tion process,which has an important impact on and improves the perfor-mance of the recommendation model.Based on the supply chain scenario of enterprise network,we propose a new algorithm of enterprise partner recommendation,which can better predict the links between manufactur-ers and suppliers in enterprise network,and recommend better parts sup-pliers for manufacturers.We find that there are higher-order connection structures in the enterprise network due to the uncertainty of the role of the enterprise in the supply chain.By statistics,we extract three kinds of higher-order connection structures as new features,and build a tensor model to describe the complex structure of the network.At the same time,because the performance of the final product in different performance indi-cators reflects the quality of the corresponding parts suppliers,we take the feedback and evaluation of consumers on the final product as the auxiliary information recommended by the enterprise partners,and use the alternat-ing direction multiplier method to decompose these data at the same time,which effectively alleviates the problem of data sparsity.In the comparison experiment on the crawled real data set,the performance is better than that on the crawled real data set General link prediction and recommendation algorithms.2.In addition to the traditional user and object entity vectors,we introduce a new entity vector that plays a positive role in the recommendation results,so as to better mine the hidden relationship between multiple entities in the real world,and improve the scope of application of the model.Combined with the classic movie recommendation problem in the recommendation system,we propose a new movie recommendation model which integrates the implicit relationship,and prove that the fusion implicit relationship has a positive effect on score prediction.We find that the members of the movie crew have a certain impact on the movie score,and the similar members of the movie crew can get a certain degree of similar feedback from the users,because the users have a subjective preference for the members of the movie crew,which can have an important impact on the prediction of the users' score for the movies that have not been scored,but there is no body in the traditional matrix decomposition method Now this implied relationship.Different from the existing methods,our model constructs a user movie crew coupling data model,fully mining the implicit relationship between users and members of the movie crew,projecting the members of the movie crew into the same implicit space,thus adjusting the user vector and movie vector in the implicit space,and better fitting the observed scoring records.The contrast experiments on the open dataset MovieLens verify the advanced nature of the algorithm.