Recommendation Algorithm Combining Multi-source Heterogeneous Side Information

The rapid development of the Internet and the continuous enrichment of information collection and sharing methods have prompted people to step into the era of big data.The rapid increase of data has brought serious "information overload" problem,and the recommendation system is proposed to solve this problem.The recommendation system analyzes the behavior records of users,excavates their personalized needs,and helps users find products that interest them but are hard to find.Nowadays,the network content is increasingly rich,the number of items is in the millions,and the user interaction data is very sparse,leading to the average effect of most association analysis algorithms.Due to the lack of information available for analysis and utilization of new users and new items,it is difficult to give accurate recommendations,which leads to the problem of cold start.At the same time,side data that can help recommendation,such as: social relational data,reviews and tags and so on,have a variety of sources,data structure and varied characteristics,is not convenient to use analysis.In addition,recommender system also faces many challenges such as improving the interpretability of the recommendation results and improving the scalability of the system.In view of the above situations and challenges,this paper adopts deep learning technology to integrate side information of users and items to enhance the recommendation performance.For different side data,different deep learning algorithms are used to analyze and extract,so as to enrich the feature representation of users and items.At the same time,mining users’ shortterm preferences,analyze users’ social relationships,building a more comprehensive user portrait,providing users with more accurate recommendation services.This paper mainly proposes two different recommendation methods for the analysis and fusion of multi-source heterogeneous side data.The main research work of this paper is as follows:1.Aiming at the problem of data sparseness and item cold start,the proposed method(MKASR,MSAKR)uses auxiliary information such as item attributes,labels,and text to organize and construct a knowledge graph,and uses knowledge graph technology to deeply mine and analyze the semantic relationship between items and items.From the perspective of semantic association,an accurate and comprehensive item feature representation is constructed,thereby improving the accuracy and divergence of recommendation results.At the same time,in terms of the integration of the knowledge graph representation learning module and the recommendation module,a multi-task learning method based on soft constraints is adopted,and the two modules are trained at the same time,so that the two tasks can complement each other,learn from each other,and promote each other,so as to improve the accuracy and generalization of the overall model.Aiming at the problem that users’ interest preferences are difficult to capture,the proposed method(MKASR)uses a bidirectional GRU network based on attention mechanism to analyze the user’s recent item interaction sequence and mine the user’s short-term preferences;The proposed method(MKASR)uses the Ripple Net algorithm to extract the relationship pairs from users to items to knowledge graph entities,and store them as in the form of triples,the knowledge graph related technology is used to analyze and mine these relationship pairs,and mine user preferences from the perspective of user and item auxiliary information association.2.Aiming at the problem of data sparseness and user cold start,the proposed model(MSAKR)analyzes graph-structured data such as user social relations,and mine trust relations between users.The proposed model uses graph convolutional neural network to extract social relations,innovatively uses graph centrality as a probability to sample neighbor nodes to filter neighbors,uses word2 vec idea to generate virtual neighbors to alleviate the sparsity of social data,and aggregate node neighbors using an attention-based approach,and build user feature vectors containing trust relationships and social associations by analyzing social records between users,thereby improving the accuracy of recommendation results.Experiments are performed on real datasets,and the proposed two recommendation algorithms outperform other benchmark models in both CTR prediction and top-K recommendation,proving that the proposed module can effectively integrate multi-source heterogeneous auxiliary information to alleviate data sparsity and cold start issues,thereby improving recommendation performance.