A deep reinforcement learning based long-term recommender system

• Published in: Knowledge-based systems 2021-02, Vol.213, p.106706, Article 106706
• Main Authors: Huang, Liwei, Fu, Mingsheng, Li, Fan, Qu, Hong, Liu, Yangjun, Chen, Wenyu
• Format: Article
• Language: English
• Subjects: Cold starts; Cold-start; Decision making; Deep learning; Deep reinforcement learning; Interactive systems; Long-term recommendation; Markov processes; Optimization; Recommender system; Recommender systems; Recurrent neural networks
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106706

Recommender systems aim to maximize the overall accuracy for long-term recommendations. However, most of the existing recommendation models adopt a static view, and ignore the fact that recommendation is a dynamic sequential decision-making process. As a result, they fail to adapt to new situations and suffer from the cold-start problem. Although sequential recommendation methods have been gaining attention recently, the objective of long-term recommendation still has not been explicitly addressed since these methods are developed for short-term prediction situations. To overcome these problems, we propose a novel top-N interactive recommender system based on deep reinforcement learning. In our model, the processes of recommendation are viewed as Markov decision processes (MDP), wherein the interactions between agent (recommender system) and environment (user) are simulated by the recurrent neural network (RNN). In addition, reinforcement learning is employed to optimize the proposed model for the purpose of maximizing long-term recommendation accuracy. Experimental results based on several benchmarks show that our model significantly outperforms previous top-N methods in terms of Hit-Rate and NDCG for the long-term recommendation, and can be applied to both cold-start and warm-start scenarios. •A novel top-N interactive recommender system based on deep reinforcement learning is proposed.•The interactions between recommender system and users are simulated by recurrent neural networks.•The proposed model can deal with both cold-start and warm-start scenarios.•Reinforcement learning and supervised learning are employed to optimize the proposed model for long-term recommendation accuracy.•Experiments and comparisons are conducted to show the merits of the proposed model.