Research map / Label-efficient and robust learning / Recognition and long-tail
Self-Paced Unified Representation Learning for Hierarchical Multi-Label 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.
Proposes SPUR, a self-paced unified representation learning framework for hierarchical multi-label classification that jointly encodes class hierarchy and instance features and orders training by level-wise loss-based hardness.
Problem
Hierarchical multi-label classification (HMLC) where existing methods use the class hierarchy only as a hard consistency constraint and uniformly sample mini-batches across levels, generalizing poorly at deep levels and with missing features.
Approach
Two lightweight encoders produce unified embeddings of instances (MLP) and of the class DAG (GIN-style node embeddings pre-trained by graph reconstruction, then parameter-free pooling); a self-paced regularizer uses level-wise losses against historical thresholds to weight instances from easy to hard.
Key results
- SPUR achieves the best AU(PRC) on 19 of 20 HMLC datasets and the best average ranking (1.10), versus 2.25 for C-HMCNN(h), the runner-up.
- On CELLCYCLE FUN, SPUR reaches 0.261 AU(PRC) versus 0.255 for C-HMCNN(h); on IMCLEF07A, 0.960 versus 0.956.
- Wilcoxon test gives p = 1.91e-6 for SPUR vs C-HMCNN(h); Friedman test p = 5.71e-17.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPUR instead of C-HMCNN(h) | HMLC models · CELLCYCLE FUN | AU(PRC): 0.255 -> 0.261 | Helped |
| Use SPUR instead of C-HMCNN(h) | HMLC models · DIATOMS | AU(PRC): 0.758 -> 0.776 | Helped |
| Use SPUR instead of HMCN-F | HMLC models · GASCH2 GO | AU(PRC): 0.465 -> 0.422 | Hurt |
| Use SPUR instead of HMCN-F | HMLC models · SEQ GO | AU(PRC): 0.447 -> 0.450 | Helped |
| Add self-paced learning to the structure-aware learner (ablation) | SPUR · FunCat tasks (Figures 4a-b) | AU(PRC) | Helped |
| Train only on hard instances rather than using the proposed self-paced strategy (ablation) | SPUR · FunCat tasks (Figures 4a-b) | AU(PRC) | Hurt |
| Add class-hierarchy DAG encoding (ablation) | SPUR-G · FunCat tasks (Figures 4c-d) | AU(PRC) | Helped |
| Use a loss-based dynamic threshold instead of K-Means difficulty estimation (ablation) | SPUR · CELLCYCLE GO and DIATOMS | AU(PRC) | Helped |
| Use a loss-based dynamic threshold instead of a linear difficulty estimator (ablation) | SPUR · CELLCYCLE GO and DIATOMS | AU(PRC) | Helped |
| Add class-instance semantic dependencies (ablation) | SPUR · CELLCYCLE FUN and SEQ FUN with missing features | AU(PRC) | Helped |
| Set λ to 0.0 instead of 0.8 (ablation) | SPUR · CELLCYCLE FUN | AU(PRC): 0.249 -> 0.260 | Helped |
| Set λ to 0.0 instead of 0.8 (ablation) | SPUR · DERISI FUN | AU(PRC): 0.190 -> 0.198 | Helped |
Surprising findings
- A smaller self-paced pace parameter (lambda = 0.0) gave the best performance in the sensitivity test, which the authors interpret as favoring a slower pace.
Limitations
- Graph-based encoder adds some computational cost relative to C-HMCNN(h) during training.
- Standard deviations omitted in the main table; only stated to be small for SPUR.
Experimental setup
Datasets: CELLCYCLE FUN DERISI FUN EISEN FUN EXPR FUN GASCH1 FUN GASCH2 FUN SEQ FUN SPO FUN CELLCYCLE GO DERISI GO EISEN GO EXPR GO
Benchmarks: CELLCYCLE FUN DERISI FUN EISEN FUN EXPR FUN GASCH1 FUN GASCH2 FUN SEQ FUN SPO FUN CELLCYCLE GO DERISI GO EISEN GO EXPR GO
Metrics: AU(PRC) average ranking training time
Baselines: MLSPL C-HMCNN(h) HMLC-LMLP CLUS-ENS HMCN-R HMCN-F
Builds on: self-paced learning graph isomorphism network (GIN) C-HMCNN max constraint loss (MCLoss) Weisfeiler-Lehman test
Closest papers by meaning
- Hierarchical Multi-Label Classification Networks (ICML 2018 · similarity 0.90)
- Modeling Label Space Interactions in Multi-label Classification using Box Embeddings (ICLR 2022 · similarity 0.88)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels (ICLR 2025 · similarity 0.86)
- Enhancing Multi-Label Classification via Dynamic Label-Order Learning (AAAI 2024 · similarity 0.86)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · 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.