Answers are grounded in this paper and cite their sources.
Label distribution learning addresses label polysemy, but acquiring precise label distributions is expensive. Label enhancement mitigates this by converting low-cost binary logical labels into continuous label distributions, reducing the reliance on fine-grained annotation. Conventional enhancement methods, however, rely on discriminative loss functions that drive scores toward discrete extremes ( or ), compromising the fidelity and continuity of recovered distributions. We propose ILE-PCA, which generates unnormalized label scores via logic-constrained target generation: for a logical , the target is drawn from ; for a logical , from . This avoids the misleading guidance of discriminative losses. To align scores with the intrinsic structure of true distributions, we apply PCA to the logical label matrix and enforce alignment between its dominant subspace and the generated scores. Experiments show that ILE-PCA excels in label distribution recovery, achieving the best average rank across datasets and outperforming all compared state-of-the-art methods in the majority of cases.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: