Commodity Recommendation Model Based On Deep Reinforcement Learning

With the expanding of the scale of the e-commerce market and increasing of the variety of products,how user can quickly and accurately find their favorite or most needed products when they faced with a wide range of products becomes the most concerned issue for users and e-commerce businesses.The product recommendation system can spontaneously find the most suitable products for users,and becomes an important method to solve the problem of “information overload”.Therefore,researching how to improve the accuracy of product recommendation is a subject of both academic value and commercial value.Traditional recommendation methods have some drawbacks such as data sparseness,cold start,and feature recognition.Deep reinforcement learning introduces deep neural network based on the reinforcement learning,which enables both automatic feature recognition ability of deep learning and strategic decision-making ability of reinforcement learning.In order to solve the problems of the traditional recommendation method,this thesis introduces the Deviling Network and the Longand Short-term Memory Network(LSTM)based on the Deep Q Network(DQN),and proposes a deep neural network structure that considers user negative feedback and commodity purchase timing.At the same time,the product recommendation model based on deep reinforcement learning is constructed to make it more suitable for usercommodity feature recognition and commodity recommendation strategy decision.The main work of this thesis is as follows:(1)Carry out a detailed review of the traditional recommendation technology and the current popular recommendation technology,summarize the shortcomings and deficiencies in the current recommendation technology,and propose the research content of this paper;(2)Analyze the shortcomings of traditional DQN network structure for commodity recommendation,and consider the deep neural network structure of user negative feedback and commodity purchase timing for the lack of design.The network uses the competition architecture(Dueling Network)and the long-term and short-term memory network(LSTM)to improve the traditional DQN network structure.To a certain extent,it solves the problem that unable distinguish the positive and negative feedback data,and the problem of unable extracting time-order feature which exist in the purchase of goods;(3)Construct a commodity recommendation model based on deep reinforcement learning.The model is based on the improved DQN network structure described above and is divided into three parts: data acquisition,offline training and online update.Firstly,the user and product related features are built in the user-commodity feature database;then the offline model is obtained by using the interactive log line between the user and the product;finally,the preprocessed data is input into the trained improved DQN network to perform the model update online;(4)Set up multiple sets of control experiments to empirically demonstrate the model,and verify the superiority of the model in the accuracy and diversity of product recommendation.The deep neural network structure considering the negative feedback of users and the timing of purchase of goods,this thesis solves the problem that the traditional DQN network can't distinguish the positive and negative feedback data,and can't extract the temporal features existing in the purchase of goods.The problem of constructing a product recommendation model based on deep reinforcement learning solves two problems in the traditional commodity recommendation model.Firstly,only consider maximizing current returns and ignoring future returns,Secondly,use the frequently occurring features in historical data to learn which make users tired.The experimental results show that the product recommendation model based on deep reinforcement learning is superior to the control model in the five groups of accuracy,recall,MAP,NDCG and commodity diversity.Based on the research content,we will continue to expand and further improve the accuracy of product recommendation.