Batch(Offline) Reinforcement Learning for Recommender System
Mohammad Amir Rezaei Gazik, Mehdy Roayaei · 2023
The explosive spread of the Internet in recent years has increased the types and amounts of big data, making it difficult for users to search for the data they need. With the continued growth of business on the Internet, e-learning, increased communication and sharing among users, and the advent of social networking, there is an undeniable need to design and implement systems that make it easier for people to search. A recommender system provides the ability to provide the most appropriate and accurate suggestions to users by checking user-related information from relevant datasets. In other words, it extracts user preferences and interests from data and makes suggestions. In this paper, we study the problem of learning recommendation systems with large datasets. We propose an offline RL framework for recommender systems to achieve high accuracy and perform recommendations quickly. Specifically, we propose a framework called Ofrec that first transforms the problem into a Markov Decision Process (MDP) and provides highly accurate and time-saving recommendations. We conduct extensive experiments on a large dataset of CIKM 2019 EComm AI and show that the proposed approach outperforms supervised learning and reinforcement learning algorithms.