Research On Enhanced Recommendation System Based On User Interest Model

With the rapid development of Internet technology,the field of e-commerce is thriving,and a large number of commercial platforms with a large number of goods have emerged,but when facing a large number of goods,it is difficult for platform users to select the goods that meet their personal interests.Therefore,there are the following problems in e-commerce platforms: on the one hand,it is difficult for users to find the goods they are interested in,and the search process takes a lot of time and energy,and the user experience is poor;on the other hand,the platform transaction volume is low,which prevents the platform transaction royalties from increasing.The recommendation system can push personalized goods for platform users through specific strategies to accelerate commodity turnover,increase transaction volume and promote parties to increase profits.Traditional recommendation systems often use collaborative filtering to get recommendation results,but the ability to fit complex features is insufficient,reduce accuracy of recommendation results.Therefore,this paper takes the goods recommendation of ecommerce platform as the research object,designs more efficient recommendation methods based on user interest extraction and reinforcement learning theory.It puts forward the clickthrough rate prediction model based on behavior delay and shared network,and the Top-N recommendation method based on reinforcement learning.In this paper,we first improve the structure of the recursive neural network,introduce the time features to determine how much user interest information should be retained at context moments.In order to improve the explanatory ability of the system,we propose a multi-layer perceptron structure with shared parameters to make the user and goods vectors to the uniform space,and design a click-through rate prediction model based on the above two points.Experiments show click-through rate prediction performance and the impact of each part of the proposed models.Secondly,on the basis of the reinforcement learning framework Actor-Critic,the underlying network is reconstructed by using the improved recursive neural network and shared network,which reduces the parameters of the model,and a Top-N recommendation list generation algorithm is proposed to convert the continuous action values of the actor output into discrete the goods IDs,and a mixed reward function is used to achieve a balance between the accuracy and diversity of the recommendation list.Based on the above implement a list-level reinforcement learning recommendation system,and prove the superiority of the proposed method on the recommendation results through experiments.