Efficient Recommender Systems via Co-Clustering-Based Collaborative Filtering

Abstract

Recommender systems became indispensable for assisting customers, users, and businesses in various domains. Collaborative Filtering (CF) is a widely used technique for generating recommendations considering user and item interactions. Many existing recommenders, such as Single Value Decomposition (SVD) and correlation, to mention a few, are based on the CF technique. These approaches suffer from two significant drawbacks. The first one is that they are computationally expensive, while the second one is the inability to cope with newly arrived user-item interactions. This leads to a situation where users’ known preferences do not change over time. However, for all practical purposes in real-time applications, there is a need to update user preferences dynamically. In this paper, we proposed a novel approach known as co-clustering-based CF that performs real-time CF considering newly arrived items, users, and ratings in rapid succession. It systematically clusters rows (users) and columns (items) with an incremental mining model. Specifically, we proposed an Efficient Co-Clustering-Based Product Recommender (ECPR) algorithm for dynamically generating recommendations that reflect the latest state of user-items-ratings dynamics. The framework is evaluated on the benchmark MovieLens dataset comprising 100,000 ratings from 943 users on 1,682 items. Comparative evaluation with existing CF methods, including SVD and Non-Negative Matrix Factorization (NNMF), demonstrates that ECPR achieves up to 3.3% improvement in Mean Absolute Error (MAE) and reduces training time by up to 60%. ECPR outperforms existing CF methods regarding computational cost and accuracy in generating recommendations.