Research And Implementation Of Commodity Recommendation System Based On Deep Reinforcement Learning

The rapid development of the Internet has accelerated the circulation of information,and the advent of the era of big data has made commodity buyers unable to handle thousands of commodities.Commodity recommendation systems need to meet the diverse needs of users in massive amounts of data.The recommendation list generated by the current product recommendation system is mostly under the premise that the user characteristics no longer change.However,the interaction between the user and the recommendation system will change the user’s shopping status.If the product recommendation system fails to record the corresponding status changes and make recommendations based on the current status,there will be a problem of failing to better reveal the characteristics of the user’s status.The research and realization of a commodity recommendation system that can reveal the characteristics of users interacting with the recommendation system will have important research and practical value.This article has two innovations.The first innovation is to propose the use of multiple elements such as product category,product brand and product ID to generate a vector containing more product information through a multi-layer deep learning network.The second innovation is to propose a state grouping mechanism,which integrates user-related product behavior design scores into the same group according to the same score.It is convenient to predict the behavior of related products based on the cosine similarity to reduce the amount of calculation,which is also speeding up the life recommendation list.Then,on the basis of the aforementioned research,this article implements the Actor-Critic framework based on deep reinforcement learning firstly.The second is to design and implement a product recommendation system that can call the Actor-Critic framework to generate a recommendation list.The core functions are divided from the three types of roles of registered users,merchants,and administrator users to realize user registration,product inquiry,product purchase,order evaluation and other common functions of product recommendation systems.The third is to store the product behavior logs of registered users in order to provide a data set for the next iteration of the recommendation model.This subject has completed the test of the core functions and verified that the product recommendation system based on deep reinforcement learning can effectively complete the functions of user registration,product query,product purchase,order evaluation,etc.And it can generate the corresponding recommendation list according to the current status of the user.The system compares the accuracy of recommendation with multiple recommendation algorithms in the short and long term.The short-term accuracy reaches 40.5%and the long-term accuracy reaches 63%.At the same time,compared with the AUC indicators of multiple models,it reaches 77.94%,which can effectively ensure the accuracy of generating recommendation lists based on user status.