Partial label learning via label-specific features enhancement and semi-disambiguation
Jinhua Hu, Jing Chai, Xianran Sun · 2024
Partial Label Learning (PLL) represents a common multi-class weakly supervised learning challenge, where each instance is linked with a set of possible labels, only one of which is the ground-truth label, and it remains unknown during the training. A common strategy for solving this multi-class problems is to decompose it into a series of binary classification problems based on certain coding rules (e.g., OvO, ECOC), and then process each binary classification problem separately using the original features. To address the deficiency of discriminative information in the original features, a feature enhancement mechanism is proposed to introduce label-specific features to improve the discriminative performance of features. Firstly, label-specific features are generated for each decomposed binary classification problem, and then fused with the original features. The fused features are used to design corresponding binary classifiers. Additionally, due to the presence of a large number of false positive labels in the training examples, a semi-disambiguation strategy can be employed to remove some false positive labels, thereby "slimming down" the candidate label set as a preprocessing step to reduce its ambiguity. The proposed partial label learning algorithm based on features enhancement and semi-disambiguation proves its effectiveness on synthetic and real-world data sets through extensive experiments.