A spectral-graph-derived weakly-supervised contrastive loss, using estimated semantic similarity as edge weights, improves noisy-label and partial-label learning.
Each runs has been repeated 3 times with different randomly-generated noise and we report the best accuracy
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Weakly-Supervised Contrastive Learning for Imprecise Class Labels
A spectral-graph-derived weakly-supervised contrastive loss, using estimated semantic similarity as edge weights, improves noisy-label and partial-label learning.