Research map / Label-efficient and robust learning / Recognition and long-tail
Calibrating Video Watch-time Predictions with Credible Prototype Alignment
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 ProWTP, a two-stage model-agnostic watch-time regression method that builds credible prototypes from the watch-ratio label distribution via HVQ-VAE and aligns instance representations to them with semi-relaxed unbalanced optimal transport plus assignment and compactness losses.
Problem
Watch-time prediction in video recommendation suffers from label-transformation issues that ignore the multimodal distribution of watch-ratio and from instance representation confusion in deep recommendation models.
Approach
Generates prototypes directly from the multimodal watch-ratio label distribution (rather than from pretrained hidden representations) using a hierarchical VQ-VAE, then aligns instance representations to these prototypes via semi-relaxed unbalanced optimal transport solved as a weighted Lasso, with assignment and compactness losses to tighten clusters.
Key results
- On WeChat, ProWTP reaches RMSE 28.47 and XAUC 0.6183 versus 28.78 and 0.6115 for the best baseline CWM.
- On KuaiRand-Pure, ProWTP reaches RMSE 40.45 and XAUC 0.7290, the best across all methods.
- Ablation shows removing SUOT or HVQ-VAE raises RMSE (e.g., 29.90 w/o SUOT vs 28.47 full on WeChat).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HVQ-VAE-generated label-distribution prototypes (ablation) | ProWTP · WeChat | RMSE: 29.12 -> 28.47 | Helped |
| Add assignment loss (ablation) | ProWTP · WeChat | RMSE: 29.45 -> 28.47 | Helped |
| Add compactness loss (ablation) | ProWTP · WeChat | RMSE: 29.38 -> 28.47 | Helped |
| Use SUOT alignment instead of direct linear prototype combination (ablation) | ProWTP · WeChat | RMSE: 29.90 -> 28.47 | Helped |
| Use ProWTP instead of traditional regression (ablation) | Watch-time prediction · WeChat | RMSE: 30.39 -> 28.47 | Helped |
| Add HVQ-VAE-generated label-distribution prototypes (ablation) | ProWTP · KuaiRand-Pure | RMSE: 41.08 -> 40.45 | Helped |
| Add assignment loss (ablation) | ProWTP · KuaiRand-Pure | RMSE: 41.35 -> 40.45 | Helped |
| Add compactness loss (ablation) | ProWTP · KuaiRand-Pure | RMSE: 41.12 -> 40.45 | Helped |
| Use SUOT alignment instead of direct linear prototype combination (ablation) | ProWTP · KuaiRand-Pure | RMSE: 42.00 -> 40.45 | Helped |
| Use ProWTP instead of traditional regression (ablation) | Watch-time prediction · KuaiRand-Pure | RMSE: 42.41 -> 40.45 | Helped |
| Replace HVQ-VAE prototype generation with VQ-VAE (ablation) | ProWTP · WeChat | RMSE: 28.47 -> 28.82 | Hurt |
| Replace HVQ-VAE prototype generation with Kmeans (ablation) | ProWTP · WeChat | RMSE: 28.47 -> 29.07 | Hurt |
Limitations
- Evaluation limited to offline metrics on three datasets with no online A/B test reported.
- Gains over the strongest baselines are modest (e.g., MAE 19.84 vs 19.90 on WeChat).
Experimental setup
Datasets: WeChat KuaiRand-Pure Short-video (private)
Benchmarks: WeChat KuaiRand-Pure Short-video
Metrics: MAE RMSE XAUC XGAUC
Baselines: Traditional Regression WLR Ordinal Regression D2Q TPM DVR CWM
Builds on: VQ-VAE semi-relaxed unbalanced optimal transport D2Q TPM weighted logistic regression
Closest papers by meaning
- CREAD: A Classification-Restoration Framework with Error Adaptive Discretization for Watch Time Prediction in Video Recommender Systems (AAAI 2024 · similarity 0.89)
- MTSTRec: Multimodal Time-Aligned Shared Token Recommender (ICML 2025 · similarity 0.83)
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation (NeurIPS 2025 · similarity 0.82)
- Causal Representation Learning for Generalisable Recommendation (NeurIPS 2026 · similarity 0.82)
- Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.82)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.81)
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.