Research map / Label-efficient and robust learning / Recognition and long-tail

Heterogeneous Label Shift: Theory and Algorithm

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Heterogeneous Label Shift (HLS) for cross-modal transfer with heterogeneous feature spaces and shifted label distributions, proposes an error decomposition theorem and a bound-minimization network (HLSAN) with reweighted adversarial alignment, evaluated on Multilingual Reuters and Wikipedia datasets.

Problem

Unsupervised knowledge transfer between source and target domains that have heterogeneous feature spaces (e.g., different modalities or languages) and shifted label distributions, where existing heterogeneous domain adaptation and label shift methods each handle only part of the problem.

Approach

An error decomposition bound separating conditional error gap and importance-weight estimation error, which motivates aligning heterogeneous latent features with a reweighted source marginal via adversarial training, plus a parallel-instance loss to bridge modalities and a regularized quadratic program for estimating label-shift weights without target labels.

Key results

• HLSAN reaches 59.9% accuracy on Reuters SP-EN Text→Text at Dirichlet γ=10, versus 25.0% for TNT and 45.1% for JMEA.
• On Wikipedia Image→Text, HLSAN achieves 81.73% at γ=2 and 85.83% at γ=10, outperforming JMEA (72.02% and 82.28%).
• Ablation on SP tasks at γ=10 shows removing the parallel-instance loss, importance weights, or domain discriminator each degrades accuracy.

Measured effects reported in the paper

Surprising findings

• Traditional label shift methods (BBSL, RLLS) perform worst, indicating heterogeneous feature spaces break standard label shift approaches.

Limitations

• Results are sensitive to hyperparameters α and β.
• Requires a small set of unlabeled parallel instances across modalities.
• Evaluated only on two cross-modal text/image classification benchmarks with simulated Dirichlet label shift.

Experimental setup

Datasets: Multilingual Reuters Collection Wikipedia

Benchmarks: Multilingual Reuters Collection Wikipedia

Metrics: accuracy

Baselines: BBSL RLLS TNT JMEA

Builds on: DANN Black-box shift learning (BBSL) Generalized label shift Heterogeneous domain adaptation (JMEA) Maximum mean discrepancy

Closest papers by meaning

• A Unified Framework for Heterogeneous Semi-supervised Learning (CVPR 2025 · similarity 0.86)
• HGDL: Heterogeneous Graph Label Distribution Learning (NeurIPS 2024 · similarity 0.86)
• Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (NeurIPS 2026 · similarity 0.85)
• Addressing Label Shift in Distributed Learning via Entropy Regularization (ICLR 2025 · similarity 0.85)
• Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation (NeurIPS 2024 · similarity 0.85)
• CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.84)