Research map / Label-efficient and robust learning / Recognition and long-tail
AttentionXML: Label Tree-based Attention-Aware Deep Model for High-Performance Extreme Multi-Label Text Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use label-specific multi-label attention instead of max-pooling (ablation) | AttentionXML · EUR-Lex, Wiki10-31K, AmazonCat-13K | P@5: 53.12 -> 59.61; 59.55 -> 66.51; 63.57 -> 66.29 | Helped |
| Use BiLSTM instead of XML-CNN's convolution and dynamic pooling (ablation) | XMTC text classification · EUR-Lex, Wiki10-31K, AmazonCat-13K | P@5: 49.21 -> 53.12; 56.21 -> 59.55; 61.40 -> 63.57 | Helped |
| Add SWA (ablation) | AttentionXML (BiLSTM+Att) · EUR-Lex, Wiki10-31K, AmazonCat-13K | P@5: 59.61 -> 61.10; 66.51 -> 68.78; 66.29 -> 66.90 | Helped |
| Increase ensemble size from 1 to 3 PLTs (ablation) | AttentionXML · Amazon-670K, Wiki-500K, Amazon-3M | P@1: 45.66 -> 47.58; 75.07 -> 76.95; 49.08 -> 50.86 | Helped |
| Increase ensemble size from 3 to 4 PLTs (ablation) | AttentionXML · Amazon-670K, Wiki-500K, Amazon-3M | P@1: 47.58 -> 48.03; 76.95 -> 77.21; 50.86 -> 51.66 | Helped |
| Increase K=M from 8 to 128 (ablation) | AttentionXML · Amazon-670K | P@1: 45.74 -> 42.96 | Hurt |
| Increase K=M from 64 to 256 (ablation) | AttentionXML · Wiki-500K | P@1: 75.07 -> 74.26 | Hurt |
| Increase PLT height from H=2 to H=4 (ablation) | AttentionXML · Amazon-670K | P@1: 45.74 -> 45.29 | Hurt |
| Add a PLT with H=2 instead of using no PLT (ablation) | AttentionXML · EUR-Lex, Wiki10-31K, AmazonCat-13K | P@5: 61.10 -> 60.88; 68.78 -> 67.27; 66.90 -> 66.28 | Hurt |
| Add a deeper PLT with H=4 instead of using no PLT (ablation) | AttentionXML · EUR-Lex, Wiki10-31K, AmazonCat-13K | P@5: 61.10 -> 60.54; 68.78 -> 65.89; 66.90 -> 65.46 | Hurt |
| Use three-PLT ensemble rather than one PLT (ablation) | AttentionXML · Amazon-670K, Wiki-500K, Amazon-3M | P@5: 36.94 -> 38.92; 44.41 -> 46.14; 43.88 -> 45.83 | Helped |
| Use AttentionXML rather than the strongest listed baseline | XMTC · All six benchmarks | P@1 and P@5: EUR-Lex P@1 83.21 -> 87.12; Wiki-500K P@5 39.68 -> 46.14 | Helped |
Closest papers by meaning
- CascadeXML: Rethinking Transformers for End-to-end Multi-resolution Training in Extreme Multi-label Classification (NeurIPS 2022 · similarity 0.92)
- LightXML: Transformer with Dynamic Negative Sampling for High-Performance Extreme Multi-label Text Classification (AAAI 2021 · similarity 0.91)
- Don't Pay Attention, PLANT It: Pretraining Attention via Learning-to-Rank (NeurIPS 2026 · similarity 0.89)
- Harnessing the Power of Beta Scoring in Deep Active Learning for Multi-Label Text Classification (AAAI 2024 · similarity 0.86)
- Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · similarity 0.85)
- ELMO : Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces (ICML 2025 · 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.