A Distributed Coordinate Descent Algorithm for Learning Factorization Machine

Abstract

Although much effort has been made to implement Factorization Machine (FM) on distributed frameworks, most of them achieve bad model performance or low efficiency. In this paper, we propose a new distributed block coordinate descent algorithm to learn FM. In addition, a distributed pre-computation mechanism incorporated with an optimized Parameter Server framework is designed to avoid the massive repetitive calculations and further reduce the communication cost. Systematically, we evaluate the proposed distributed algorithm on three different genres of datasets for prediction. The experimental results show that the proposed algorithm achieves significantly better performance (3.8%–6.0% RMSE) than the state-of-the-art baselines, and also achieves a 4.6–12.3 ✕ speedup when reaching a comparable performance.