Research map / Label-efficient and robust learning / Recognition and long-tail
Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation
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.
Extends RKME learnware specifications with pseudo-labeled marginal and conditional distributions to learn a label-aware unified subspace and improve identification for heterogeneous-feature learnwares.
Problem
Reusing high-performing models trained on heterogeneous feature spaces in a learnware dock system without raw data or auxiliary co-occurrence data, when user tasks lack a model with matching feature space.
Approach
Extend the RKME specification to a labeled reduced set (features, pseudo-labels from model outputs) that encodes both marginal and class-conditional distributions; use it to drive supervised+contrastive+reconstruction subspace learning across feature blocks and to match users by both marginal and conditional distances.
Key results
- On 23 classification datasets, Ourcls averages 0.674 accuracy vs 0.599 for TabPFN and 0.593 for Hetero; Ourunify averages 0.661.
- On 7 regression datasets, Ourunify averages RMSE 0.194 vs 0.215 for LightGBM and 0.223 for Hetero.
- Ensemble of learnware with self-training outperforms self-training with 100 labeled points in most cases, and still helps in about 21% of classification cases with 5000 labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the specialized classification specification instead of the unified specification (ablation) | Heterogeneous learnware identification and reuse · Classification tasks | accuracy: 0.661 -> 0.674 | Helped |
| Replace the RKME specification with the proposed label-aware specification (ablation) | Heterogeneous learnware identification and reuse · Classification tasks | accuracy: 0.652 -> 0.661 | Helped |
| Replace the RKME specification with the proposed label-aware specification (ablation) | Heterogeneous learnware identification and reuse · Regression tasks | RMSE: 0.196 -> 0.194 | Helped |
| Add conditional-distribution encoding to the classification specification (ablation) | Ourunify · Classification tasks | accuracy: 0.661 -> 0.674 | Helped |
| Ensemble learnware predictions with a self-training model | User tasks with limited labeled data · Classification tasks | accuracy | Helped |
| Ensemble learnware predictions with a self-training model | User tasks with limited labeled data · Regression tasks | RMSE | Helped |
Surprising findings
- Alignunlabel (heterogeneous alignment without labels) performs worse than plain self-training on user tasks.
- Learnware benefit persists even with 5000 labeled user samples for some datasets, such as kin8nm.
Limitations
- Requirement generation for regression uses a self-trained model for pseudo-labeling and specification for classification only in the specialized mechanism.
- Experiments limited to tabular data with LightGBM developer models and 100 labeled user samples in main results.
Experimental setup
Datasets: Tabzilla benchmark datasets (30)
Benchmarks: Tabzilla (30 datasets: 23 classification, 7 regression)
Metrics: accuracy RMSE average rank win/tie/loss
Baselines: LightGBM TabPFN Alignunlabel Alignlabel Transtab Xtab Hetero Ourbasic
Builds on: RKME specification learnware paradigm Self-VPCL Hetero Beimingwu
Closest papers by meaning
- Learnware Specification via Label-Aware Neural Embedding (AAAI 2025 · similarity 0.89)
- Towards Making Learnware Specification and Market Evolvable (AAAI 2024 · similarity 0.86)
- Learnware Specification via Dual Alignment (ICML 2025 · similarity 0.85)
- Heterogeneous Label Shift: Theory and Algorithm (ICML 2025 · similarity 0.85)
- Any-Way Meta Learning (AAAI 2024 · similarity 0.84)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.84)
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.