A reinforcement learning recommender system using bi-clustering and Markov Decision Process

Abstract

Collaborative filtering (CF) recommender systems are static in nature and does not adapt well with changing user preferences. User preferences may change after interaction with a system or after buying a product. Conventional CF clustering algorithms only identifies the distribution of patterns and hidden correlations globally. However, the impossibility of discovering local patterns by these algorithms, headed to the popularization of bi-clustering algorithms. Bi-clustering algorithms can analyze all dataset dimensions simultaneously and consequently, discover local patterns that deliver a better understanding of the underlying hidden correlations. In this paper, we modelled the recommendation problem as a sequential decision-making problem using Markov Decision Processes (MDP). To perform state representation for MDP, we first converted user-item votings matrix to a binary matrix. Then we performed bi-clustering on this binary matrix to determine a subset of similar rows and columns. A bi-cluster merging algorithm is designed to merge similar and overlapping bi-clusters. These bi-clusters are then mapped to a squared grid (SG). RL is applied on this SG to determine best policy to give recommendation to users. Start state is determined using Improved Triangle Similarity (ITR similarity measure. Reward function is computed as grid state overlapping in terms of users and items in current and prospective next state. A thorough comparative analysis was conducted, encompassing a diverse array of methodologies, including RL-based, pure Collaborative Filtering (CF), and clustering methods. The results demonstrate that our proposed method outperforms its competitors in terms of precision, recall, and optimal policy learning.