Efficient and Scalable Matrix Factorization Transfer with Review Helpfulness for Massive Data Processing

Abstract

We explore the sparsity problem associated with recommendation system through the concept of transfer learning (TL) which are normally caused by missing and noisy ratings and or review helpfulness. TL is a machine learning (ML) method which aims to extract knowledge gained in a source task/domain and use it to facilitate the learning of a target predictive function in a different domain. The creation and transfer of knowledge are a basis for competitive advantage. One of the challenges prevailing in this era of big data is scalable algorithms that process the massive data in reducing computational complexity. In the RS field, one of the inherent problems researchers always try to solve is data sparsity. The data associated with rating scores and helpfulness of review scores are always sparse presenting sparsity problems in recommendation systems (RSs). Meanwhile, review helpfulness votes helps facilitate consumer purchase decision-making processes. We use online review helpfulness votes as an auxiliary in formation source and design a matrix transfer framework to address the sparsity problem. We model our Homogenous Fusion Transfer Learning approach based on Matrix Factorization HMT with review helpfulness to solve sparsity problem of recommender systems and to enhance predictive performance within the same domain. Our experiments show that, our framework Efficient Matrix Transfer Learning (HMT) is scalable, computationally less expensive and solves the sparsity problem of recommendations in the e-commerce industry.