Research map / Label-efficient and robust learning / Recognition and long-tail
Predicting Label Distribution from Ternary Labels
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.
Proposes predicting label distributions from ternary (+1/0/-1) annotations and a CateMO categorical distribution that models ternary labels as a function of label description degree, usable as a loss or probability model in existing label enhancement methods.
Problem
Acquiring ground-truth label distributions for label distribution learning is expensive, and binary multi-label annotations force experts to make definitive judgments in ambiguous cases, reducing annotation efficiency and accuracy.
Approach
Introduce a three-way annotation scheme (1, 0 uncertain, -1) and a CateMO categorical distribution with probabilistic monotonicity and orderliness constraints that maps label description degrees to ternary labels, plugged into existing label enhancement methods as a loss or generative likelihood.
Key results
- CateMO-based variants (e.g., GL-CateMO, LR-CateMO, MR-CateMO) outperform DT, MSE and LL baselines across datasets and are close to the ground-truth label distribution training.
- Theorem 1 shows ternary labels have lower expected approximation error than binary labels when (k-3/4)^2 + (t-1/4)^2 <= 3/8 and t <= k, with averaged errors of 7/36-type expressions favoring ternary in most cases.
- Annotation-cost analysis indicates ternary labels achieve better prediction performance than binary labels while requiring less annotation time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the MSE loss with CateMO for ternary-label enhancement | GL · JAFFE | KL: 0.162 ± 0.016 -> 0.158 ± 0.016 | Helped |
| Replace the MSE loss with CateMO for ternary-label enhancement | GL · Painting | KL: 0.571 ± 0.027 -> 0.542 ± 0.022 | Helped |
| Replace the MSE loss with CateMO for ternary-label enhancement | GL · Music | KL: 0.104 ± 0.007 -> 0.103 ± 0.007 | Helped |
Surprising findings
- Binary labels are superior to ternary labels only in extreme cases of the predefined thresholds, which are rarely used in practice.
Limitations
- CateMO precision parameters are pre-fixed rather than learned, because the constraints in Theorem 2 are interdependent and hinder gradient-based optimization.
- Datasets are small (213 to 360 instances) and the ternary labels were manually re-annotated by three experts with majority voting.
- Data and code are not released due to dataset licensing.
Experimental setup
Datasets: JAFFE Painting Music
Benchmarks: JAFFE Painting Music
Metrics: Chebyshev distance KL divergence Cosine Intersection Spearman's rho
Baselines: GL-DT GL-MSE LR-DT LR-MSE MR-DT MR-MSE GL-LL LR-LL MR-LL Ground-Truth
Builds on: label enhancement GL LR MR LDL-LRR
Closest papers by meaning
- Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.89)
- Generative Calibration of Inaccurate Annotation for Label Distribution Learning (AAAI 2024 · similarity 0.89)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.87)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.87)
- Enhancing Multi-Label Classification via Dynamic Label-Order Learning (AAAI 2024 · similarity 0.86)
- Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.