Research map / Video / Video segmentation and tracking
CREAD: A Classification-Restoration Framework with Error Adaptive Discretization for Watch Time Prediction in Video Recommender Systems
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 CREAD, a classification-restoration framework that predicts watch time via multiple binary classifiers over discretized buckets, with an error-adaptive discretization (EAD) balancing learning and restoration errors.
Problem
Watch time prediction in video recommender systems is hard because the target is highly imbalanced and long-tailed; existing discretization schemes (equal-width, equal-frequency) trade off learning error against restoration error.
Approach
Recasts watch time regression as multiple ordinal binary classifiers, restores the expectation from their probabilities, and chooses bucket boundaries via a calibrated CDF-based function that minimizes a theoretical bound combining learning error and restoration error.
Key results
- On KuaiRec, CREAD reaches XAUC 0.6009 and MAE 3.2150, versus D2Q at 0.5677 XAUC and 3.3090 MAE.
- On the Industrial dataset, CREAD gives MAE 18.6716 versus 19.0740 for D2Q, about 2.11% lower.
- Online A/B test on Kwai App over 5 days: watch time increased by 0.291% on average with 20% traffic.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CREAD instead of OR | Watch-time prediction · KuaiRec | XAUC: 0.5883 -> 0.6009 | Helped |
| Use CREAD instead of OR | Watch-time prediction · KuaiRec | MAE: 3.2946 -> 3.2150 | Helped |
| Use CREAD instead of D2Q | Watch-time prediction · CIKM | XAUC: 0.6568 -> 0.6671 | Helped |
| Use CREAD instead of OR | Watch-time prediction · CIKM | MAE: 1.9157 -> 1.7928 | Helped |
| Use CREAD instead of D2Q | Watch-time prediction · Industrial | XAUC: 0.6423 -> 0.6459 | Helped |
| Use CREAD instead of D2Q | Watch-time prediction · Industrial | MAE: 19.0740 -> 18.6716 | Helped |
| Replace EAD with equal-frequency discretization (ablation) | CREAD · KuaiRec | MAE: 3.2150 -> 3.2477 | Hurt |
| Replace EAD with equal-width discretization (ablation) | CREAD · KuaiRec | MAE: 3.2150 -> 3.2492 | Hurt |
| Set EAD beta to 3.0 instead of 0.5 (ablation) | CREAD · KuaiRec | XAUC: 0.5977 -> 0.6009 | Helped |
| Set EAD beta to 3.0 instead of 0.5 (ablation) | CREAD · KuaiRec | MAE: 3.2419 -> 3.2150 | Helped |
| Add ordinal regularization (ablation) | CREAD · KuaiRec | MAE: 0.0132 in MAE | Helped |
| Add ordinal regularization (ablation) | CREAD · KuaiRec | XAUC: 0.33 in XAUC | Helped |
Surprising findings
- Equal-frequency discretization beats equal-width on long-tailed data in the KuaiRec ablation, despite the theory suggesting trade-offs.
- Ordinal regularization degrades performance in the first 12 buckets, which the authors call a sacrifice for overall generality.
Limitations
- The authors did not provide official code for baselines, so they reimplemented them.
- The theoretical analysis assumes tabular inputs and smooth distributions; beta must be tuned as a hyperparameter because C_w and C_b cannot be obtained theoretically.
- Industrial dataset and online results are not reproducible by external researchers.
Experimental setup
Datasets: KuaiRec CIKM Industrial (Kuaishou streaming short-video logs)
Benchmarks: KuaiRec CIKM Industrial
Metrics: MAE XAUC
Baselines: VR WLR OR D2Q
Builds on: D2Q Weighted Logistic Regression Ordinal Regression
Closest papers by meaning
- Calibrating Video Watch-time Predictions with Credible Prototype Alignment (ICML 2025 · similarity 0.89)
- Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.83)
- Causal Representation Learning for Generalisable Recommendation (NeurIPS 2026 · similarity 0.82)
- Influential Exemplar Replay for Incremental Learning in Recommender Systems (AAAI 2024 · similarity 0.82)
- MRMLREC: A Two-Stage Approach for Addressing Data Sparsity in MOOC Video Recommendation (Student Abstract) (AAAI 2024 · similarity 0.82)
- Exploiting Fine-Grained Skip Behaviors for Micro-Video Recommendation (AAAI 2025 · 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.