Binary Classification From $M$-Tuple Similarity-Confidence Data

Abstract

A recent advancement in weakly-supervised learning utilizes pairwise similarity-confidence (Sconf) data, allowing the training of binary classifiers using unlabeled data pairs with confidence scores indicating similarity. However, extending this approach to handle high-order tuple data (e.g., triplets, quadruplets, quintuplets) with similarity-confidence scores presents significant challenges. To address these issues, this paper introduces M-tuple similarity-confidence (Msconf) learning, a novel framework that extends Sconf learning to $M$-tuples of varying sizes. The proposed method includes a detailed process for generating $M$-tuple similarity-confidence data and deriving an unbiased risk estimator to train classifiers effectively. Additionally, risk correction models are implemented to reduce potential overfitting, and a theoretical generalization bound is established. Extensive experiments demonstrate the practical effectiveness and robustness of the proposed Msconf learning framework.