Research On Recommendation Algorithm With Multi-context Information

With the development of information technology and the popularization of mobile terminals,Internet data continues to expand.Massive information enriches people's lives,but it is also difficult for some people to locate the data they need.Problems to be solved.The recommendation system pays attention to the user's interests and hobbies based on the user's historical data,helps users find the information they need,and can alleviate the information explosion problem to a certain extent.Recommendation algorithms have received continuous attention from academia and industry,and different types of technologies have been proposed one after another.One type of representative algorithm is collaborative filtering.The collaborative filtering algorithm is simple and efficient,and mainly discovers the user's interest based on the user's interaction data.However,when the interaction data between the user and the product is insufficient,the effect of recommendation is not dominant.In real systems,users and commodities often have rich contextual information including text,pictures,and tags,which can provide more available resources for the recommendation system.However,this information presents diversity and complexity in type and structure,and how to effectively integrate such multicontext information is one of the important challenges faced by current recommendation algorithms.To this end,this article focuses on two different types of contextual information: 1)For discrete label data,using users,products,and labels as nodes,construct a "user-item-tag" heterogeneous network.Define a variety of meta-paths with semantics in the heterogeneous graph,and use the meta-path to walk upstream in the heterogeneous graph in a direction,and convert the structural information of the graph into data in the form of a sequence.Meta-path can capture the high-order correlation of users(or items)in the sense of discrete attributes,and use SPPMI to measure this correlation.2)For continuous text information,the feature of each node is represented by text information,and the similarity of text features between nodes is measured through Laplace regularization constraints,which can supplement the semantic relevance between users and users,products and products.Finally,different similarity networks are constructed according to the characteristics of different types of contexts,and the objective function is designed to perform joint matrix decomposition under the constraints of multiple context information networks,and to learn the characterization of users and commodities.We have conducted sufficient experiments on multiple data sets,and the results show that the collaborative filtering algorithm that integrates multiple types of contextual information not only improves the recommendation effect and meets the personalized needs of users,but also this method can effectively alleviate the problem of data sparsity.