A Review of Label Disambiguation and Feature Selection in Partial Multilabel Learning

Partial multilabel learning addresses ambiguous supervision in which each instance is assigned a candidate label set containing its relevant labels and additional false positives. Feature selection in this setting faces a circular dependency: reliable labels are needed to identify informative features, while informative features are needed to distinguish relevant labels from misleading candidates. This review examines that dependency through a focused synthesis of recent partial multilabel feature selection methods and related work on disambiguation, partial label learning, and label enhancement. We organize the literature along two complementary dimensions: the evidence used to resolve label ambiguity and the mechanism through which the resulting supervision controls feature selection. Eight representative feature selection methods illustrate latent alignment, mutual information reconstruction, dual-space fusion, multiscale label distributions, sparse regression, fuzzy relevance, and neighborhood rough sets. The comparison shows that similar terminology can conceal different supervision constraints, feature scoring rules, and optimization dependencies. We therefore distinguish label confidence from label description distributions and original feature selection from representation learning. We also examine differences in noise generation, feature budgets, downstream classifiers, and validation protocols that limit comparisons across published results. The review identifies uncertainty calibration, minority-label preservation, reliable structural assumptions, and reproducible evaluation as priorities for developing feature selection methods under ambiguous multilabel supervision.