Product Recommendation And Enterprise Partner Link Prediction Algorithm Based On Coupled Tensor Decomposition

The main purpose of a recommendation system is to recommend information of interest to users according to their characteristics of interest and historical behavior,and to help them make decisions quickly and effectively.The recommendation system is widely used in the field of electronic commerce.Besides,link prediction in social networking relationships is also one of the research directions in recommendation systems.In this thesis,we base ourselves on the field of electronic commerce,and study how to recommend personalized automobile products to users and how to recommend high-quality parts suppliers to automobile manufacturers,namely the link prediction between the enterprise.In the production and sale of automobiles,there are supply chain and demand chain.Supply chain refers to the supply of parts by automobile suppliers to manufacturers,and the manufacturer produces automobile sales to users.The demand chain refers to the feedback evaluation given to the manufacturer and supplier after the user buys the car.The traditional method of product recommendation usually depends on the user's historical evaluation of the product or the attribute information of the user and the product.However,the information only involves two dimensions:the user and the product.The useful data extracted is sparse and difficult to achieve good recommendation results.Enterprise partner link prediction mainly depends on the attribute of enterprise nodes and network topology,and predicts the future development direction of the network according to the existing supply relations among the nodes.However,the existing link prediction methods usually only consider the low-order network structure in the network,ignoring the influence of high-order connection patterns on the evolution of network structure.In addition,the sparsity of link data in enterprise relationship network also makes it difficult to obtain accurate link prediction results.The high-order connection patterns in enterprise networks and evaluation information contained in supply and demand chains can compensate for the sparse availability of data in link predictions and product recommendations.The automotive product recommendation and the enterprise partner link prediction that this article studies do not treat the two as separate issues.Instead,we explore the high-order network structure in the enterprise network,and consider the matching degree and mutual influence between user requirements and auto manufacturers.Besides,the effective information in the supply chain is extracted to help to get more accurate product recommendation and link prediction results.Therefore,this article has designed a product recommendation and a enterprise partner link prediction model that integrates the supply-demand chain and high-order network structure.The main work is summarized as follows:1.This thesis develops an automobile product recommendation algorithm that integrates supply chain information.In order to make a more comprehensive study of personalized product recommendation,the users' scores of evaluation indexes of the cars they have bought have been made to form a sparse three-order tensor.Extract supply relationship between different enterprises and production relationship between manufacturer and automobile from supply chain,to form auxiliary information matrices and construct the coupling data model of the auxiliary information matrix and the scoring tensor,which reflects the influence of the supply chain information on the user's choice of products.2.This thesis develops a prediction model of enterprise partner link that integrates demand chain and high-order network structure.We study the high-order connection patterns in the enterprise network,mine and count the number of triads formed between enterprises,and construct tensors formed by different types of triads.Combining the evaluation information of the users and the affiliation between the automobile and the manufacturer extracted from the demand chain,a matrix-tensor coupling data model is constructed to solve the sparseness problem of linked data.Then we study the evolution of the enterprise network structure.3.For the coupled data model formed by product recommendation and enterprise partner link prediction,the alternating direction multiplication method(ADMM)is used to solve the problem,which improves the accuracy and speed of model decomposition.We conduct a large-scale experiment on real data captured from car homes and Chinese automobile suppliers online.The experimental results show that our proposed model can obtain better results than traditional algorithms.