GBRIP: Granular Ball Representation for Imbalanced Partial Label Learning

TL;DR

GBRIP addresses imbalanced partial label learning by introducing a coarse-grained granular-ball (GB) representation coupled with a GB-based graph and a multi-center loss (MCL). Through unsupervised 2NN clustering, the feature space is partitioned into roughly balanced GBs, enabling more accurate inter- and intra-class balance modeling and robust label disambiguation via a refined label confidence matrix. The joint objective combines $\mathcal{L}_{ce}$, a center-based loss $\mathcal{L}_{mc}$, and a prior-regularization term $\mathcal{L}_{pr}$ to mitigate the impact of outliers and imbalance, with w and p updated in an alternating fashion and class priors estimated by a moving-average scheme. Experiments on CIFAR10-LT, CIFAR100-LT, real-world PLL datasets, and SUN397 demonstrate that GBRIP consistently outperforms state-of-the-art PLL methods across multiple imbalance and ambiguity settings, highlighting its practical impact for robust learning under challenging label noise and imbalance.

Abstract

Partial label learning (PLL) is a complicated weakly supervised multi-classification task compounded by class imbalance. Currently, existing methods only rely on inter-class pseudo-labeling from inter-class features, often overlooking the significant impact of the intra-class imbalanced features combined with the inter-class. To address these limitations, we introduce Granular Ball Representation for Imbalanced PLL (GBRIP), a novel framework for imbalanced PLL. GBRIP utilizes coarse-grained granular ball representation and multi-center loss to construct a granular ball-based nfeature space through unsupervised learning, effectively capturing the feature distribution within each class. GBRIP mitigates the impact of confusing features by systematically refining label disambiguation and estimating imbalance distributions. The novel multi-center loss function enhances learning by emphasizing the relationships between samples and their respective centers within the granular balls. Extensive experiments on standard benchmarks demonstrate that GBRIP outperforms existing state-of-the-art methods, offering a robust solution to the challenges of imbalanced PLL.

Figures (3)

• Figure 1: Practical example of a real-world long-tailed PLL dataset CIFAR100-LT. There are both intra-class imbalanced samples for the majority candidate labels "Airport" and the minority candidate labels "Forest". Different from the existing methods that ignore the impact of intra-class imbalance, our proposed GBRIP can consider both inter-class and intra-class imbalance.
• Figure 2: The basic idea of GBRIP: Using 2NN unsupervised clustering, imbalanced samples between and within classes are gradually divided into smaller coarse-grained GB spaces, forming a new GB-graph representation.
• Figure 3: Performance comparisons on the SUN397 dataset with $\psi=0.05$.