Research On Recommender System Based On Deep Reinforcement Learning

Recommender systems are widely used in various fields of the Internet to solve the problem of information overload in the current era.It can not only help users quickly filter out the content of interest,but also creates huge commercial value for Internet companies.Analyzing users is the foundation and pre-work of the personalized recommender system.Based on the analysis of a certain user,it can be found that the user is interested in different items.Users,items,and other ancillary information,such as label information and item types,together form a heterogeneous information network.In general,the recommendation algorithm will directly describe the user as an explicit or implicit rating,but in doing so,it ignores the interconnection between users,items,and other entity nodes in the information network.In order to solve this problem,this paper will make useful recommendations based on the mining of heterogeneous information networks.Previously,many methods for solving recommendation tasks in heterogeneous information networks were based on meta-paths.However,many existing methods completely rely on experts to provide meta-paths,and few work discusses how to generate high-quality meta-paths.This has two drawbacks.One is that in a large and complex heterogeneous information network,manual retrieval of meta-paths may be very cumbersome and difficult.Second,the meta-path provided by experts may be artificially biased.Therefore,I propose a method based on deep reinforcement learning to mine meta-paths in heterogeneous information networks.Previously,reinforcement learning has been used in many recommender systems,but different from existing methods,this paper defines the environment of reinforcement learning as a heterogeneous information network composed of users,objects,and other information.In general,this paper proposes a recommender system algorithm in heterogenous information networks based on deep reinforcement user analysis.The specific research content is as follows:1.The environment is modeled as a heterogeneous information network,which is composed of users,objects,and different information sources.Between users and unobserved items on the network,this paper uses the following multiple iteration training process.There is a meta-path library that retains the meta-paths generated during each iteration.Initially,the meta-path library is empty and filled with meta-paths given by experts.Then,the meta-path tried by the reinforcement learning agent in each iteration is added to the meta-path library.The updated meta-path library is used to train the reinforcement learning agent in the next iteration.Repeat this process until no new knowledge can be acquired or the maximum number of iterations is reached.2.A meta-path-based user analysis method is proposed,which records the weight of each user’s potential preference for items.After the meta-path-based user analysis process,a user-based collaborative filtering method is proposed to complete the Top-N recommendation task.3.Experiments were conducted on three data sets.The proposed method is compared with the baseline model on the scoring prediction task and the Top-N recommendation task,this paper also analyzes the quality of the generated meta-paths,experimental results demonstrate the effectiveness of the proposed method.