Balance learning to rank in big data

Abstract

We propose a distributed learning to rank method, and demonstrate its effectiveness in web-scale image retrieval. With the increasing amount of data, it is not applicable to train a centralized ranking model for any large scale learning problems. In distributed learning, the discrepancy between the training subsets and the whole when building the models are non-trivial but overlooked in the previous work. In this paper, we firstly include a cost factor to boosting algorithms to balance the individual models toward the whole data. Then, we propose to decompose the original algorithm to multiple layers, and their aggregation forms a superior ranker which can be easily scaled up to billions of images. The extensive experiments show the proposed method outperforms the straightforward aggregation of boosting algorithms.