Using a large number of real-life datasets from The Washington Post, we explore three novel ideas in recommender systems: (1) how to learn individual preference differences deduced from different rating distributions of users; (2) how to improve the sparsity of a user-item rating matrix; and (3) how to handle different time decay patterns of popularity of items across users and domains. Then, based on the separate improvements, we propose to develop a unifying solution integrating three dimensions of ideas that could improve recommendation results. Our results will not only have significant intellectual merit in integrating three important understudied issues in recommendation, but also have broader impact on many e-commerce applications that employ some types of recommendation at their cores. Further, we will attempt to implement the developed solutions in the standard SQL to maximize its utility in other research and education.

The three novel ideas are relatively understudied separately and never studied together. While the data sparsity issue, for instance, has been known and around for years, few attempts exist to improve the sparsity of an input rating matrix directly (thereby become a method-agnostic improvement of sparsity of a rating matrix). Similarly, to our best knowledge, few existing works attempted to exploit personal preferences, deduced from different individual rating distributions. Finally, no works exist to use varying time decay models across items, users, and domains in a holistic and principled fashion. Therefore, the success of this project will bear a huge significance in recommender system research.