Reproduce “GBRIP: Granular Ball Representation for Imbalanced Partial Label Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jintao Huang, Yiu-ming Cheung, Chi-man Vong, Wenbin Qian

GBRIP tackles partial-label learning, where examples have multiple candidate labels but one hidden correct label. It groups feature-similar examples into granular balls without true labels, using their structure to estimate correct labels and pull features toward ball centers. On synthetic CIFAR10-LT, accuracy is 83.97 ± 1.66%, versus HTC’s 77.71 ± 1.12%, at ψ = 0.5 and γ = 200; this finding is limited to that setting.

Key insights

  • Class imbalance in partial-label learning has both an across-class dimension and a within-class feature-distribution dimension; addressing only class counts may leave label estimates unreliable.
  • Coarse feature groups can provide useful structure for label disambiguation even when exact training labels are unavailable.
  • Ablations show that performance relies on the broader combination of grouping, multi-center loss, and training components; removing both granular-ball representation and multi-center loss sharply reduced accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.