Research Area: Machine Learning
The increasing popularity of the recommender system deeply influences our decisions on the Internet, which is a typical continuous interaction process between the system and its users. Most previous recommender systems heavily focus on optimizing recommendation accuracy while neglecting the other important aspects of recommendation quality, such as diversity of recommendation list. In this study, we propose a novel recommendation framework to optimize the recommendation list for the Top-N task, named Collaborative Filtering-based Deep Reinforcement Learning (CFDRL), which promotes the diversity of recommendation results without sacrificing the recommendation accuracy. More specifically, to effectively capture the continuous user-item interaction for recommendations, we adopt the deep reinforcement learning (DRL) to update the recommendation strategy dynamically according to the user-s real-time feedback. Meanwhile, to generate diverse and complementary items for recommendation, we design a diversity-aware reward function that can lead to maximizing reward with the trade-off between diversity and accuracy. Besides, to alleviate the disadvantage of DQN that directly picking the recommendations with the highest Q-values from the unselected items, we define a modified greedy explore policy with jointly CF model. It firstly utilizes CF model to sort the items and divide them into two part according to the item similarity, then with a probability the agent selects from them and generates an action list with the modified greedy explore policy. The experimental results conducted on two real-world e-commerce datasets demonstrate the effectiveness of the proposed model.
Keywords:
Author(s) Name: Tao Wang Xiaoyu Shi, Mingsheng Shang
Journal name: Big Data
Conferrence name:
Publisher name: Springer
DOI: https://doi.org/10.1007/978-981-16-0705-9_16
Volume Information:
