Research Of Recommendation Algorithm Based On Joint Similarity And Generative Adversarial Nets

The rapid-developed Internet technology provides rich data support for recommendation system,but make the dataset sparser as well.Dataset sparsity results in some problems,such as low recommendation accuracy,easy over fitting of training model and sensitivity to noise,etc.Therefore,how to improve the recommendation performance without changing the sparsity of dataset is important for recommendation system.In the view of the problems mentioned above,the main research of this thesis is as follows:(1)To deal with low recommendation accuracy in sparse dataset,a joint similarity combined with forgetting function algorithm is proposed.First,we discuss similar relationships among users,analyze characteristics of existing similarity measurement methods.For each similarity relationship between users,chose the corresponding similarity measurement method and combined.Second,we introduce forgetting function based on joint similarity so as to solve the characteristics of user interest changing with time.Experimental results show that the proposed algorithm has higher recommendation accuracy than the compared algorithms.(2)As for training model easy over fitting and sensitivity to noise,we propose an adversarial training-based mean Bayesian personalized ranking algorithm.First,we discuss Bayesian personalized ranking algorithm,analyze the shortcomings of positive and negative feedback division of interaction between users and items by Bayesian personalized ranking algorithm with a result of a mean Bayesian personalized ranking algorithm,which is more consistent with the characteristics of user behavior.Second,to improve the anti-disturbance capability of the algorithm,this paper introduces adversarial training of generative adversarial networks into mean Bayesian personalized ranking algorithm.Experimental results demonstrate that the proposed algorithm has better recommendation performance than the classical compared algorithms.