New Method Improves Rank Estimation with Noisy Data

A novel learning framework, Stochastic Order Learning (SOL), has been introduced to address the challenging problem of rank estimation when dealing with noisy ordinal labels. Instead of treating noisy labels as simple corruption, this approach reinterprets the task as a stochastic ordering problem, acknowledging that each data instance can be associated with multiple plausible ranks rather than a single definitive one. SOL is designed to effectively capture the inherent uncertainty within ordinal label data. It achieves this by learning an embedding space through the combination of two distinct objectives: a discriminative loss that structures interactions between instances and their centroids, and a stochastic order loss that enforces probabilistic relationships between the ranks of different instances. Extensive testing on diverse datasets has shown that SOL significantly improves the reliability of rank estimation, even in the presence of various forms and levels of label noise.