Research On Reinforcement Learning Recommendation Algorithm Based On User Evaluation

With the rapid development of computer technology,the world has became a sea of information.People are faced with more and more choices while enjoying the convenient services brought by massive information.How to mine the characteristics of users' interests and make accurate and personalized recommendations for users has became a research focus in this thesis.The collaborative filtering recommendation algorithm in the current mainstream personalized recommendation algorithm is the most effective and widely used algorithm.The collaborative filtering algorithm predicts the preference of the target user by calculating the similarity between users or items and predicts users' evaluation of the item according to the neighboring user.But on the one hand,the collaborative filtering algorithm has problems such as cold start and matrix sparsity;on the other hand,the current recommendation algorithm only considers the user's score of the item but ignores the users' personalized evaluation of the item.so it is difficult to fully exploit the users' interest characteristics.There will be no personalized recommendations.In view of the above problems in the recommendation algorithm,this thesis take advantage of reinforcement learning and adding user evaluation factors to realize personalized recommendation algorithm based on user evaluation.The main work contents are as follows:A collaborative filtering recommendation algorithm based on value function estimation is proposed.Aiming at the cold start and matrix sparsity problems of collaborative filtering recommendation algorithm,this thesis Introduces the idea of value function estimation in reinforcement learning and proposes a method to calculate the similarity between users by comparing the state value function between users to replace the previous similarity calculation method to alleviate the problem of matrix sparsity.By controling the weight convergence speed to solve the cold start problem.Reinforcement learning evaluates the user's current state by learning the user's state value function,and updates the weights in the user state value function by the gradient descent method,and then uses the updated weights to calculate the new state value function,and iterates until it reaches the end.status.Finally,the collaborative recommendation is performed according to the neighboring users with high similarity with the target user state value function.A reinforcement learning recommendation algorithm based on user evaluation is proposed.Faced with the problem in the current recommendation algorithm that only the user score is difficult to dig deep into the problem and the user interest feature lacks personalized recommendation.The user evaluation factor is added on the basis of the reinforcement learning recommendation algorithm.On the one hand,user evaluation vector is added to the value function estimation of the state value function.The vector describes the frequency that the user evaluationthe used by the target user of all users to reflect the degree of discrimination between the target user and the entire user group,The state value function formed by the evaluation,weight and scoring can better estimate the state of the user;on the other hand,in the update of the weight,the user evaluation is added to evaluate the internal frequency of the user in each user.,thereby further enhancing the convergence effect of the state value function.Through the above methods,the algorithm can deeply mine user interest characteristics for personalized recommendation.The above two algorithms are tested on the Movielens dataset,and compared with the current mainstream collaborative filtering recommendation algorithm and the existing classical reinforcement learning recommendation algorithm.Experiments show that the above two algorithms can effectively alleviate the cold start and matrix sparsity problems of collaborative filtering recommendation algorithm.At the same time,the enhanced learning recommendation algorithm based on user evaluation can deeply explore user interest characteristics and improve recommendation accuracy.