Improving Automatic Recommendation By Modeling Intrinsic And Extrinsic Motivation With Q-learning

Personalized recommendation systems can provide users with objects or services that play a vital role in mitigating information overload issues.In the real scene,the real reason that influences the user's choice is the user's psychological motivation.According to the explanation of motivation in psychology,it mainly includes two parts: intrinsic motivation and extrinsic motivation.Among them,the intrinsic motivation is based on the internal thinking of the individual,and the extrinsic motivation is derived from the influence of the others.In order to achieve joint modeling of dual motivation,we chose a combination of deep learning and reinforcement learning.First,we use the deep learning model,Stacked Denoising Auto-Encoder(SDAE)model,to use the user's historical selection information and the potential feature representation of the user and the object to simultaneously model the user and the object and get the preliminary user and the potential feature representation of the object.The Bayesian Probability Matrix Decomposition(BPMF)is used to optimize the potential feature representation of users and objects.Based on the potential feature representation of users and objects,we propose Q-learning(Off-policy)to model the two psychological motives of users.In the specific modeling,we mainly have three aspects of improvement and contribution:(1)the combination of deep learning model and BPMF model,it is more efficient and alleviates the problem of data sparseness and cold start,and reduces the over-fitting caused by many parameters of the deep learning model,which improves the accuracy of the model.(2)The fusion of two psychological motives based on Q-learning(Off-policy)further alleviates the cold start problem and makes the recommendation performance significantly improved.In order to test the validity of our proposed model,we conducted experiments on two open source movie datasets and tested them with two evaluation indicators.The experiment performed very well.