Research Of Recommender System Technology Based On Reinforcement Learning For Debiasing Problem

In recent years,with the continuous deepening of global digital transformation,all aspects of people’s daily lives have been innovated and optimized by Internet technology,and the speed of information production and dissemination has reached the pinnacle of human history.In the digital process,recommendation system is one of the essential and important productivity.Excellent recommendation systems can improve the efficiency of people’s acceptance of information,help users dig deeper into their interests,bring users a better user experience,and help content providers to gain more revenue.However,to achieve personalized recommendation effects of "one thousand people,one thousand faces",recommendation systems need to solve various problems,among which the bias problem is a very worthy direction to explore.Due to its obscurity,bias is harming personalized recommendation goals in an imperceptible way.Therefore,this thesis deeply study the popularity bias problem and the source and mechanism of exposure bias and combines reinforcement learning for recommendation algorithms,proposes a reinforcement learning for recommendation algorithm that eliminates the corresponding bias impact,and verifies its effectiveness on multiple datasets.The main research contents are as follows:(1)Design a Popularity Separate(PS)reinforcement learning algorithm framework based on structural causal inference,and construct PS-DQN and PS-REINFORCE models by combining two classic reinforcement learning for recommendation algorithms,DQN and REINFORCE.By studying the causal relationship between users’ interaction tendencies and popularity factors of items,this thesis identifies the popularity factor as a confounding factor and uses the backdoor adjustment method to eliminate the influence of the confounding factor,thereby alleviating popularity bias.The algorithm’s recommendation performance and popularity recommendation tendency were tested and compared on three different-sized real datasets,and the results proved the algorithm’s effectiveness.(2)Propose the User Based Reward Shift Deep Q-Learning(URS-DQN)algorithm.This thesis studies the mechanism of exposure bias and proposes a user-oriented Reward Shift method to address the inaccuracy of negative feedback in interaction data.After clustering users,the interaction tendencies of similar users are transferred,and the classic reinforcement learning algorithm DQN is combined to propose the URS-DQN algorithm to mitigate the impact of exposure bias and extract users’ true interests more accurately.In the experimental verification,this thesis compared and analyzed the algorithm’s effectiveness with four baseline algorithms on two datasets,and the experimental results showed that the recommendation performance and training efficiency of the URS-DQN algorithm were better.In addition,the effectiveness of each module of the algorithm was explored and verified through multiple ablation experiments.