Research And Implementation Of Recommendation System Based On Deep Reinforcement Learning

Nowadays,the recommendation system has become a powerful tool for users when dealing with information overload problems.The recommendation system can analyze the characteristics between users and recommended items,or through existing users and behavioral record between items,which helps users to make better choice in the future.Although traditional machine learning has achieved remarkable success in the field of recommendation systems,there are still many deficiencies that are difficult to solve or overcome,such as limited representation capabilities of the model,and it is not well adapted to dynamic user interest changes.In recent years,the flourishing development of deep learning has brought new vitality to the recommendation system.At the same time,since Alpha Go,deep reinforcement learning has been greatly developed in the past two years,not only in the field of games,but also in areas such as control and natural language processing.In the scenario where most recommendation systems work,the process of continuous interaction and feedback between users and users is very similar to the environment in which reinforcement learning is good at processing.In summary,this thesis combines deep learning reinforcement learning as well as recommendation system and proposes a recommendation system model based on deep reinforcement learning.the main work includes:1)Combining the recommendation system with reinforcement learning,Markov modeling the process of the interaction between user and recommendation system,defining the state transition matrix and reward function in the process in detail,and constructing a dynamic long-term interaction recommended system environment;2)Using the gated recurrent unit in deep learning and introducing item embedding,the recommended items are low-latitude coded so that the state represented by the recommended item history has a richer information representation.3)Combining the classic value-based method Deep Q Network in reinforcement learning with the interactive recommendation system,a deep reinforcement learning recommendation system based on Top N problem modeling is proposed,and proposes various methods based on difficult training and strong data relevance in actual training.Combining the advantages of the reinforcement learning valuebased method and the policy-gradient method,a deep reinforcement learning recommendation model based on the Actor-Critic framework is proposed,which greatly improves the recommendation on large-scale data.4)In the aspect of activation function,the defects of Re LU function on this problem are discussed,and the Dice activation function based on data characteristics is introduced,which greatly improves the performance of some models under the same conditions,and analyzes the time of the new activation function and space cost situation.5)We test all the deep reinforcement learning method and with some classic populary methods in nowadays recommendation system.The test including cold start and warm start processes and differences in recommended performance in longterm tracking of user dynamic preferences,providing direction for follow-up research.