Hub for Outliers, Spokes for Inliers: Uniform Latent Space Construction for Dual-Mismatched Semi-Supervised Learning

The gist The proposed method constructs a hub-spoke latent geometry to address dual-mismatched semi-supervised learning by uniformly distributing known classes around a central hub and guiding high-uncertainty unknown class samples toward this hub, which induces a low-evidence region for modeling unknown classes.

Problem Formulation and Motivation

Semi-supervised learning typically assumes that labeled and unlabeled data share an identical class distribution and label space, but this setting is often violated as unlabeled data may be imbalanced and contain unknown class samples causing mismatches in both class distribution and label space. This dual mismatch leads to majority classes dominating the latent space and unknown class samples being overconfidently misclassified, degrading feature discriminability and pseudolabel quality. Existing methods lack an effective strategy to handle these realistic and challenging settings and therefore suffer from significant performance degradation.

Hub-Spoke Geometry Construction

The proposed solution involves constructing a hub-spoke geometry where known classes are uniformly distributed around a central hub and each class forms compact clusters around its prototype, while the hub provides an anchor for a low-evidence region specifically designed for high-uncertainty unknown class samples. This geometry promotes inter-class separation and intra-class compactness, mitigating representation bias caused by class imbalance. The global structure regularization loss Lg integrates an inter-class separation loss Lig and a prototype geometry loss Log to enforce a globally uniform feature distribution where all known classes are uniformly distributed around a central hub.

Class-Adaptive Re-weighting

To enhance feature discriminability for minority classes, the method employs a class-adaptive re-weighting strategy that dynamically prioritizes samples from minority classes. For each sample, the weight wi is defined as 1 + max 0, qi · ∆yi∥qi∥2 ·∥∆yi∥2, where qi is the displacement from the overall class prototype to the sample feature representation and ∆yi is the overall deviation vector of the class. Since minority classes exhibit a significantly larger deviation between the overall class prototype and the labeled class prototype, samples belonging to minority classes receive higher weights, adaptively shifting the model’s focus toward these classes.

Unknown Class Identification and Modeling

High-uncertainty unknown class samples are identified using an evidence-based classifier g(·) which produces an evidence vector e, and predictive uncertainty is derived from the Dirichlet strength S = PC c=1 αc, with the uncertainty defined as u = C/S. During training, samples are ranked by their uncertainty u in descending order to select the top-k most uncertain samples as candidates from unknown classes. To model these candidates, an unknown class sample clustering loss Loc is introduced to explicitly model their feature distribution by pulling them toward the hub, which induces a low-evidence region around the hub and Theorem 1 shows that high uncertainty among the selected candidates and strong hub-centered clustering yield a quantitatively low-evidence hub region and correspondingly high evidential uncertainty.

Training Process and Results

The training process is structured into three phases: a warm-up phase focusing on learning stable initial feature representations using Lwarm1, a second warm-up phase integrating the supervised and unsupervised loss guided by Lwarm2, and finally, the main training phase leveraging the full objective function Loverall. Extensive experiments on CIFAR-10, CIFAR-100, SVHN, Food-101, and ImageNet-100 datasets demonstrate that our method achieves new state-of-the-art (SOTA) performance gains by up to 3.25% across various settings. The hub-spoke geometry construction (HGC) module substantially improves minority class accuracy, e.g., from 33% to 64% in the lb4 setting on CIFAR-10. Furthermore, the unknown class identification and modeling (UCIM) module boosts pseudo-label accuracy from 47% to 60% in the same lb4 setting on CIFAR-10. The method achieves significant improvements over all LTSSL methods in 75% of cases and over all OSSL methods in 76.7% of cases.

Ablation and Robustness

Ablation studies confirm that progressively integrating the modules leads to consistent performance gains, with the full model achieving the best performance. The method is robust to the choice of k, as both k = 10% and k = 20% yield nearly identical performance and significantly outperform the k = 0% baseline across all three settings. The superior performance, particularly in the lb10 setting, highlights the advantage of our hub-spoke geometry, which leverages evidence-based uncertainty to guide high-uncertainty unknown class samples toward a shared low-evidence hub. The method maintains relatively stable performance across all settings despite substantial changes in the diversity of unknown classes.

Conclusion

In this work, we addressed the challenging and realistic scenario of dual-mismatched semi-supervised learning, where unlabeled data exhibits an imbalanced class distribution and contains unknown class samples. Our proposed method integrates a structured hub-spoke geometry with evidence-based unknown class identification. The hub-spoke geometry ensures a discriminative and robust feature representation space, while the evidence-based classifier provides well-calibrated uncertainty estimates. Extensive experiments demonstrate that our method consistently outperforms previous state-of-the-art methods under various dual mismatch conditions.

--- Page 9 ---

The gist The proposed method constructs a hub-spoke latent geometry to address dual-mismatched semi-supervised learning by uniformly distributing known classes around a central hub and guiding high-uncertainty unknown class samples toward this hub, which induces a low-evidence region for modeling unknown classes.

How it works

The proposed method addresses the dual mismatch by constructing a discriminative feature representation space via a hub-spoke geometry. This geometry is achieved through two types of regularization: global structure regularization (Lg) and local structure regularization (Ll). The global structure loss Lg integrates an inter-class separation loss Lig and a prototype geometry loss Log to enforce a globally uniform feature distribution where all known classes are uniformly distributed around a central hub.

Unknown Class Modeling

To effectively identify high-uncertainty samples from unknown classes, an evidence-based classifier g(·) is employed for uncertainty quantification, deriving uncertainty u = C/S where S = PC c=1 αc. High-uncertainty unknown class samples are selected by ranking them by their uncertainty u in descending order to form the set DˆO,k. To model these candidates, an unknown class sample clustering loss Loc is introduced to explicitly model their feature distribution by pulling them toward the hub. This design induces a low-evidence region around the hub, indicating the absence of known-class evidence.

Training Process

The training process is structured into three phases to ensure stable learning. The main training phase leverages the full objective function Loverall = Lsup + Lunsup + λ (Lecon + Lg + Ll + Ledl + Loc). The class-adaptive weight wi from Eq. 9 is incorporated into the Lunsup term to adaptively focus on minority classes.

1. Improved feature representation space via Hub-Spoke Geometry: The proposed geometry structures features such that known classes are uniformly distributed around a central hub and each class forms compact clusters around its prototype, which is shown to promote inter-class separation and intra-class compactness, mitigating representation bias caused by class imbalance.

2. Enhanced uncertainty handling for unknown classes: The system can identify high-uncertainty samples by leveraging an evidence-based classifier where The predictive uncertainty is derived from the Dirichlet strength S = PCc=1 αc, with the uncertainty defined as u = C/S, and then guides these samples toward the hub, which induces a low-evidence region around the hub.

3. Adaptive minority class focus: The model dynamically prioritizes minority classes by defining a weight for each sample based on its deviation from the overall prototype: the weight wi for sample xi is defined as wi = 1 + max 0, qi · ∆yi∥qi∥2 ·∥∆yi∥2, ensuring samples belonging to minority classes receive higher weights, thereby adaptively shifting the model’s focus toward these classes.

4. Robust unknown class modeling: By pulling identified high-uncertainty samples toward the hub, the system ensures that these samples exhibit low evidence by construction, which improves the separation between known and unknown classes in the feature space and their uncertainty distributions.

Semi-supervised learning typically assumes that labeled and unlabeled data share an identical class distribution and label space. However, this setting is often violated: unlabeled data may be imbalanced and contain unknown class samples, causing mismatches in both class distribution and label space. Such dual mismatch leads to majority classes dominating the latent space and unknown class samples being overconfidently misclassified, degrading feature discriminability and pseudo-label quality. To address this, we propose a hub-spoke latent geometry, where known classes are uniformly distributed around a central hub and each class forms compact clusters around its prototype, while the hub provides an anchor for a low-evidence region specifically designed for high-uncertainty unknown class samples. Integrated with an evidence-based classifier, this geometry ultimately enhances feature discriminability and uncertainty separation by mitigating majority-class domination through structured feature organization and guiding high-uncertainty unknown class samples toward the hub. Extensive experiments show that our method outperforms state-of-the-art methods, with a maximum improvement of 3.25% across various settings.