Information Recommendation System Integrated With Social Network And Based On Deep Reinforcement Learning

As one of the most effective information filtering systems,recommendation system plays an important role in the era of information overload.A number of techniques for personalized recommendation have been proposed,however,most techniques still suffer from following two limitation:(1)the recommendation process is considered to be static and the effect of dynamic characteristics in the recommendation system is ignored.(2)the data sparsity issue has not been resolved.In order to mitigate above two problems,two recommendation models are proposed in this paper: Recommendation System based on Deep Reinforcement Learning and Attention Mechanism(DRA)and recommendation model DRA+ which combines model DRA and social information.Recommendation model DRA.The interaction between users and recommendation system is regarded as a sequential decision making process,and the deep reinforcement learning model DDPG is used as the basic model.Considering the definition of reinforcement learning in the recommendation system,the method of PMF is used to pre-train the item vector to constitute the input.Meanwhile,the user state will be reasonable expressed by explicitly combining user vector with item vectors.In addition,the attention mechanism is designed by using the items' location information and the similarity between items and action to further explore the dynamic of users' interest.Extensive experiments on three real-world datasets MovieLens(100k),MovieLens(1M)and Jester(2)are conducted.The results demonstrate the model DRA can improve recommendation performance by mining user preference.Recommendation model DRA+.On the basis of model DRA,social network information is additionally integrated to alleviate the impact of data sparsity on recommendation performance.There are two major information extracted from socialinformation: trust relationship and similarity relationship.In order to make full use of these two kinds of information,each relationship is divided into two part,explicit and implicit trust relationship and global and local similarity relationship.Then an adaptive weight adjustment method which can better simulate user's propensity on choosing items is proposed.Finally,extensive experiments on two real-world datatsets FilmTrust and CiaoDVD is conducted.The results prove that the impact of data sparsity on recommendation performance can be alleviated by using social network information.