Research On Recommendation System Fused With Multi-source Heterogeneous Auxiliary Information

The recommendation system is very important to solve the problem of information overload,assist users to find items that meet their needs,and achieve supply and demand balance.However,traditional collaborative filtering recommendation algorithms face data sparsity and cold start problems.When there is little user-item interaction data,the accuracy of the algorithm drops dramatically.In addition,new users and new items in the system cannot be recommended due to lack of preference data.In this regard,it is generally recognized as an effective solution to integrate user item auxiliary information for recommendation.Generally speaking,auxiliary information comes from a wide range of sources and is heterogeneous.Fusion of these multi-source heterogeneous auxiliary information for recommendation has challenges such as insufficient utilization of auxiliary information,poor scalability,and mechanization of fusion methods.In view of the above problems,this paper focuses on the representation and fusion of multi-source heterogeneous auxiliary information,and proposes two different recommendation methods.The main research work of this paper is as follows:1.A multi-layer perceptron-based recommendation model for multi-source heterogeneous auxiliary information is proposed.The ability of neural networks to process multi-source heterogeneous data is used to treat the knowledge graph as a special type of auxiliary information.The perceptron machine learns the latent representation of the corresponding entity of the item,and combines the item structure information and semantic information to expand the latent representation vector of the item to improve the collaborative filtering algorithm.The model uses multi-task learning technology,utilizes knowledge graph embedding tasks to assist recommendation tasks,and designs a joint training method to optimize the parameters of two tasks in the model at the same time.The effectiveness of the proposed model in CTR prediction and Top-N recommendation tasks is verified on the real-world dataset.2.A recommendation architecture based on attention mechanism that fuses multi-source heterogeneous auxiliary information is introduced.The attention mechanism is introduced to characterize the different impacts of different types of auxiliary information on the final recommendation task,and the feature vector of different items is linearly weighted.This way of fusion improves the traditional fusion or summation feature fusion method,realizes the extraction of useful information and the suppression of noise information.Considering the user's more fine-grained interest in the item,a specific multi-level attention recommendation model is designed for attributes and textual auxiliary information.Word-level and text-level attention structures are used to measure user interest at different granularities.The real-world dataset was used to verify the performance of the model on the rating prediction task.