Reinforcement learning -based for recommender system

Abstract

The widespread of E-commerce over the years leads to the dramatic growth of customers and products and poses some significant hurdles for recommender systems. It is required to develop new recommender system technologies that can swiftly generate high-quality recommendations, especially for very large-scale problems. Over the past few years, there has been significant progress in recommendation methodologies, encompassing conventional recommendation methods like collaborative filtering, content-based recommendation, and matrix factorization, as well as deep learning-based technologies. Due to its capacity to recognize non-linear user-item relationships and work with several data sources, including images and text, deep learning in particular offers substantial advantages in completing challenging assignments and handling complicated data. As a result, it is being used in recommender systems more often. To maximize customer transactions, reward signals from the environment can be employed in reinforcement learning to learn and optimize their decisions. However, training a recommender system is often difficult because it is necessary to expose users to irrelevant recommendations. As a result, learning the policy through logged implicit feedback is vital, which is difficult due to the lack of off-policy settings and negative rewards (feedback). This thesis is based on the fundamental concept of self-supervised reinforcement learning in the context of sequential recommendation tasks. The objective of our research is to investigate the utilization of reinforcement learning in recommendation systems, drawing upon the methodologies employed by prominent companies such as JD, Microsoft, ByteDance, Alibaba, and Google in their respective recommendation systems. We endeavor to implement the Top-K method, which is one of Google's renowned techniques for proposing videos on the YouTube platform. Initially, the algorithm is re-executed using an e-commerce dataset to comprehend its functionality, as well as its input and output. Subsequently, we endeavor to reapply the aforementioned approach to the MovieLens dataset to evaluate its outcomes. Ultimately, a resolution can be reached regarding a recommender system that utilizes reinforcement learning.