GreedyBoost: An accurate, efficient and flexible ensemble method for B2B recommendations

Abstract

Recommender systems have achieved great success in finding relevant products and services for individual customers, e.g. in B2C markets, during recent years. However, due to the diversity of enterprise clients' requirements it is still an open question on how to successfully apply existing recommendation techniques in the B2B domain. This paper presents GreedyBoost - an accurate, efficient and flexible ensemble method for product and service recommendations in the B2B domain. Given a set of base models, GreedyBoost can sequentially add base models to the ensemble by a linear approach to minimize training error, so that the ensemble process is efficient. Meanwhile, GreedyBoost does not have any special requirement on base models and evaluation metrics, so that any kind of client requirements and sale & distribution purposes can be adapted. Experimental results on real-world B2B data demonstrate that GreedyBoost can achieve higher recommendation accuracy compared with two popular ensemble methods.