Research On Recommendation Algorithm Based On Multi-task Learning And Knowledge Graph

With the rapid penetration of the Internet and the widespread dissemination of information,we are in an era full of data.There are 1 billion Internet users in China,and more information is being produced every second.This rapid data accumulation has also led to excessive dissemination of information,and many consumers have fallen into an endless torrent of information.With the advancement of technology,recommender systems have emerged,which can not only help users discover valuable information,but also present the information to those who are interested in it.The recommendation system has been applied in many fields in real life,especially the personalized recommendation has more and more research and development projects and landing practice,but there are still some challenges,such as the lack of analyzable information for new users and new projects,resulting in The selection result is not accurate enough,resulting in the cold start problem.The number of items is hundreds or even thousands,and the interaction information between users and items is relatively rare.The traditional recommendation algorithm can not solve the problem of data sparsity elegantly.Using knowledge graphs as auxiliary information to generate recommendations can not only alleviate the recommendation problem,but also make personalized recommendations more accurately,and the results of the recommendations are also interpretable.By importing the relationship information between users and items,the knowledge graph can effectively mitigate the cold start and data sparse problems of the recommendation management system.However,the work of integrating knowledge graphs and recommendation systems faces many challenges.At present,most of them use embedding methods for integration,that is,transforming each entity and relationship into a vector representation of a certain dimension,but there is a problem that the training target does not match the downstream tasks.The pathbased method also has its problems.The artificial design of meta-paths often requires certain expert advice,and the effectiveness and comprehensiveness are difficult to guarantee.The graph embedding task can be used to assist the recommendation task by adopting a multi-task framework,but how these two tasks interact to make the recommendation result more accurate is a hot topic of concern.This paper proposes an improved multi-task learning architecture,which cleverly combines the recommendation module and the knowledge graph module,and fuses knowledge through the knowledge graph entity and relational features to make the recommendation results more interpretable.Research the initialization vector method of the multi-task learning framework model,explore the interaction between high-level and low-level features,and find a better representation method for the feature vectors of users and items.By embedding the knowledge map into the module assisted recommendation module,the feature vectors of the two modules are cross compressed,and their respective potential features are automatically shared.This enables learning about high-dimensional feature interactions between items and knowledge map entities in the recommendation module.The attention mechanism is used to aggregate user and item features,and at the same time capture the correlation between users and social relations,so as to mine more user feature information.Through multiple experiments on real data sets,the result of proposed recommendation model is more accurate than other benchmark models.The recommendation performance is improved by cleverly integrating the multi-head attention mechanism,and the problems of cold start and data sparseness are improved.