Learning from Noisy pairwise dissimilarities and unlabeled data
Shuying Huang, Junpeng Li, Changchun Hua, Yana Yang · 2024
Machine learning from weak supervision has become a hot topic recently, particularly in classification using similar/dissimilar data and unlabeled data.Learning pairwise similar and dissimilar data enhances data security.However, in real-world scenarios, pairwise dissimilarity data often gets contaminated with pairwise similarity data.To overcome this bottleneck, this paper proposes a model for learning from noisy pairwise dissimilarities and unlabeled (nDU) data, aiming to address the issue of pairwise similarity data being mixed into pairwise dissimilarity data.The proposed method establishes an empirical risk function for noisy pairwise dissimilarity data, provides theoretical guarantees through estimation error bounds, and validates the approach through experiments.This novel weakly supervised learning algorithm effectively mitigates the impact of noise in pairwise dissimilar samples, resulting in an effective binary classifier.