Research map / Label-efficient and robust learning / Knowledge distillation
Cross-Task Knowledge Distillation in Multi-Task Recommendation
Chenxiao Yang, Junwei Pan, Xiaofeng Gao, Tingyu Jiang, Dapeng Liu, Guihai Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add auxiliary fine-grained ranking relations (ablation) | CrossDistil · TikTok TaskA | Multi-AUC: 0.6510 -> 0.7411 | Improved |
| add auxiliary fine-grained ranking relations (ablation) | CrossDistil · TikTok TaskB | Multi-AUC: 0.8005 -> 0.8341 | Improved |
| calibrate teacher predictions before distillation (ablation) | CrossDistil · TikTok TaskA | Multi-AUC: 0.7396 -> 0.7411 | Improved |
| calibrate teacher predictions before distillation (ablation) | CrossDistil · TikTok TaskB | Multi-AUC: 0.8312 -> 0.8341 | Improved |
| apply teacher error correction (ablation) | CrossDistil · TikTok TaskA | Multi-AUC: 0.7399 -> 0.7411 | Improved |
| apply teacher error correction (ablation) | CrossDistil · TikTok TaskB | Multi-AUC: 0.8310 -> 0.8341 | Improved |
| transfer fine-grained knowledge from auxiliary ranking teachers rather than same-task regression teachers (ablation) | CrossDistil · TikTok TaskA | Multi-AUC: 0.6901 -> 0.7411 | Improved |
| transfer fine-grained knowledge from auxiliary ranking teachers rather than same-task regression teachers (ablation) | CrossDistil · TikTok TaskB | Multi-AUC: 0.8014 -> 0.8341 | Improved |
| avoid direct cross-task distillation by using conflict-filtering auxiliary ranking tasks (ablation) | CrossDistil · TikTok TaskA | AUC: 0.7269 -> 0.7494 | Improved |
| avoid direct cross-task distillation by using conflict-filtering auxiliary ranking tasks (ablation) | CrossDistil · TikTok TaskB | AUC: 0.9184 -> 0.9513 | Improved |
| use calibration (ablation) | CrossDistil · TikTok TaskA | LogLoss: 0.5832 to 0.5703 | Improved |
| use calibration (ablation) | CrossDistil · TikTok TaskB | LogLoss: 0.0623 to 0.0337 | Improved |
Closest papers by meaning
- STEM: Unleashing the Power of Embeddings for Multi-Task Recommendation (AAAI 2024 · similarity 0.89)
- Direct Routing Gradient (DRGrad): A Personalized Information Surgery for Multi-Task Learning (MTL) Recommendations (AAAI 2025 · similarity 0.86)
- Active Large Language Model-Based Knowledge Distillation for Session-Based Recommendation (AAAI 2025 · similarity 0.85)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.85)
- Bagging-Expert Network for Multi-Task Learning: A Depolarization Solution in Multi-Gate Mixture-of-Experts (AAAI 2025 · similarity 0.85)
- Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 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.
