A New Cascade-hybrid Recommender System approach tailored for the Retail Market

Abstract

Recommender systems ought to increase profit from product sales, by carefully recommending selected items to users. For this, recommendations need to be relevant, novel and diverse. There are many approaches to this problem, each with its own advantages and shortcomings. This paper proposes a novel way to combine model, memory and content-based approaches in a cascade-hybrid system, where each step refines the previous one, sequentially. In addition to this, a straightforward way to easily incorporate time-awareness into rating matrices is also advanced. This hybrid system focuses on being intuitive, flexible, robust, auditable and avoid heavy performance costs. Evaluation metrics such as Novelty Score are also formalized and computed, in conjunction with Catalog Coverage and mean recommendation price to better capture the recommender’s performance.