Research On The Method Of Deep Reinforcement Learning Recommendation Based On Knowledge Graph

With the rapid development of deep learning,the combination of deep learning and recommendation system has become a hot research direction of recommendation system.The traditional recommendation system based on deep learning has the following problems: Static recommendation,only considering short-term return,weak exploratory ability.The emergence of deep reinforcement learning recommendation provides a new solution to the above problems by using the special learning mechanism of reinforcement learning,and injects new vitality into the research of recommendation system.The research on reinforcement learning recommendation is still at the initial stage,and there are still some problems such as no connection between modeling projects,unstable training,slow convergence,etc..The knowledge graph,as a semantic network that reveals the relationship between entities,can introduce richer relational information for each project,therefore,it is meaningful to introduce knowledge map into reinforcement learning recommendation to mine the deep relationship between items and make more accurate and effective recommendation.The main work of this paper as follows:(1)The Markov decision process in the recommended scenario is designed.In order to make the reward value more reflect the user’s real feedback,the acquisition of positive and negative feedback and the reward function are designed by making full use of the user’s interactive information.Gated recurrent unit is used to deal with the embedding of interactive items,generate state representation,and model the state transition mode.(2)A multi task recommendation model based Reinforcement Learning with Knowledge representation Enhanced(KERL4Rec)is proposed.The model associates reinforcement learning recommendation and knowledge representation learning through feature combination sharing unit,provides better recommendation strategy for recommendation system through alternating learning,and improves training speed.(3)The Markov decision process adaptability setting in the KERL4 Rec model is analyzed,and two reward design variants of piecewise linear and gradient reward design are proposed.Two state modeling schemes of positive and negative feedback interactive state and user interactive state are also proposed.(4)The multi task reinforcement learning recommendation model proposed in this paper is compared with the classical recommendation model and the classical reinforcement learning recommendation model.The accuracy and cumulative return under different recommendation steps are taken as the evaluation indexes of recommendation effect.The performance of multi task reinforcement learning recommendation model under different reward function design and different state function design is compared.The experimental results show that the multi task reinforcement learning recommendation model proposed in this paper can better model the user’s dynamic behavior,has higher recommendation accuracy and average return in continuous multi-step recommendation,and also significantly improves the training efficiency,which proves the effectiveness of the model.