User Behavior Prediction In Recommender System Based On Deep Learning And Knowledge Graph

In recent years,as deep learning has demonstrated powerful representation capabilities in the fields of voice,image,and text,researchers have begun to apply it to recommender system,that is,to predict the next user-item interaction based on current interaction records and features.At present,the application of deep learning in recommender systems has achieved great success.However,mainstream recommendation models based on deep learning are faced with the problem of sparse input vectors,insufficient modeling of low-order features,and insufficient information in recommendation modules when modeling features.In allusion to the problems mentioned above,the main research contents are as follows:Firstly,a recommendation model based on two-stage deep learning(TMR)is proposed to solve the problem of sparse input vectors and insufficient modeling of low-order features.The input vectors are embedded to dense vectors by the pre-training stage,and the initial representation of the vector is optimized to reduce the impact of the sparse input vector on the model performance and efficiency.Then,Deep FM is used to extract features and combine low-order feature interaction with high-order feature interaction.The experiments show that the recommendation model based on two-stage deep learning solves the problem of sparse input vectors to a certain extent,and the introduction of low-order feature makes the representation of feature vectors more accurate,which achieves a good effect on the Movie Lens-1M data set.Secondly,a recommendation model of knowledge graph embedding based on the random walk(RWRM)is proposed to solve the problem of insufficient information caused by sparse interaction records.The node sequence was sampled by random walk,and the entity was modeled based on the neighbor nodes,which made full use of the homogeneity and isomorphism of knowledge graph.The semantic matching of the triples is performed through the neural association model,and learned the final embedding representation for the entity,and it is trained alternately with the recommendation module.The experiments show that the recommendation model of knowledge graph embedding based on the random walk makes full use of the features of neighbor nodes in the knowledge graph,making the entity embedding vector representation more accurate.In the case of sparse data sets,this model can better model user-item interactions and predict user behaviors.Finally,a recommendation model based on knowledge graphs sharing information(ISRS)is proposed to solve the problem of insufficient utilization of information in the knowledge graph.In this paper,when combining the knowledge graph with the recommender system,the interest preference of different users in the item is fully considered.The similarity between the user vector and the item vector expresses whether the user has interest in the item,and the interest of the user and the neighbor nodes is used as the weight to model the current entity.By designing the information sharing unit,the information in the knowledge graph is fully introduced into the recommendation module to supplement the item information.The experiments show that the recommendation model based on knowledge graph sharing information proposed in this paper performs well in CTR prediction and Top-K recommendation,and the model also shows better performance in movie recommendation,book recommendation and music recommendation.