Stochastic Order Learning: An Approach to Rank Estimation Using Noisy Data

Chaewon Lee, Seon-Ho Lee, Chang-Su Kim

Keywords: rank estimation, label noise, order learning

Abstract: A novel algorithm, called stochastic order learning (SOL), for reliable rank estimation in the presence of label noise is proposed in this paper. For noise-robust rank estimation, we first represent label errors as random variables. We then formulate a desideratum that encourages reducing the dissimilarity of an instance from its stochastically related centroids. Based on this desideratum, we develop two loss functions: discriminative loss and stochastic order loss. Employing these two losses, we train a network to construct an embedding space in which instances are arranged according to their ranks. Also, after teaching the network, we identify outliers likely to have extreme label errors and relabel them for data refinement. Extensive experiments on various datasets show that the proposed SOL algorithm yields decent rank estimation results even when labels are corrupted by noise.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 12490