Deep learning-based learning to rank with ties for image re-ranking
Piplong Zhao, Ou Wu, Liyuan Guo, Weiming Hu, Jinfeng Yang · 2016
In existing learning to rank problems, the learned ranking function sorts objects according to their predicted scores. Therefore, a full-ordering object list is obtained even if two or more objects have almost identical degrees of relevance (or called objects with ties). For objects containing ties, a more reasonable ranking approach is to learn a ranking function which can judge both the preference and ties relationships among objects. In this paper, we propose a new pairwise ranking algorithm and apply it to image re-ranking. Specifically, we utilize deep learning to re-rank images based on a new loss function. The ties-relationship is considered in both training and testing process. As a result, the learned ranking function can be used to rank objects containing ties. The experimental results demonstrate the effectiveness of the proposed algorithm.