Introduction
In modern short-video platforms such as TikTok and Kwai, predicting user "watch time" is critical for optimizing recommendation algorithms. Unlike simple click-through rate (CTR) prediction, watch time serves as a proxy for user satisfaction and engagement. However, accurately predicting this metric is challenging due to the extreme right-skewed distribution of the data—where most videos are watched for only a few seconds, while a tiny fraction are watched to completion—and the inherent difficulty in optimizing for ranking rather than just absolute magnitude.
The CREAD Framework
To address these challenges, the authors introduce CREAD (Classification-Restoration with Error-Adaptive-Discretization). This framework shifts the problem from direct regression to a series of ordinal classification tasks, followed by a mathematical restoration of the continuous value.
The framework operates in three distinct stages:
- Discretization: Converting continuous watch time into discrete buckets using a novel error-adaptive approach.
- Classification: Employing $M$ binary classifiers to predict the probability that a user's watch time exceeds specific thresholds ($P(y > t_m)$).
- Restoration: Reconstructing the final continuous prediction by integrating these survival probabilities, leveraging the property of expectation: $E[y] = \int P(y > t) dt$.
Error-Adaptive Discretization (EAD)
The core innovation of CREAD is the Error-Adaptive Discretization (EAD) module. Traditional discretization methods often fall into two traps:
- Equal-Width Discretization: Creates many empty buckets in the long tail, leading to high learning error due to sparse data.
- Equal-Frequency Discretization: Results in extremely wide buckets in the long tail, leading to high restoration error because the exact value within the bucket is unknown.
EAD mathematically balances these two competing errors. By using a calibrating function, it finds an optimal middle ground that ensures enough samples per bucket for stable learning while keeping bucket widths narrow enough for precise restoration.
Training Objectives and Regularization
CREAD is trained using a multi-objective loss function that combines standard classification cross-entropy with a global restoration loss (Huber loss). A key component is Ordinal Regularization, which enforces temporal logic: the probability of watching a video for a longer duration must be less than or equal to the probability of watching it for a shorter duration. This constraint is particularly effective for improving predictions in the difficult-to-model long-tail buckets.
Key Results and Industrial Impact
The framework was rigorously tested on both public benchmarks (KuaiRec, CIKM) and a massive industrial dataset from the Kwai App, involving over 6 billion impressions.
- Offline Performance: CREAD significantly outperformed state-of-the-art methods like D2Q. On the KuaiRec dataset, it achieved a 5.85% improvement in XAUC, a specialized ranking metric for watch time.
- Online Impact: In a large-scale A/B test, CREAD delivered a 0.291% increase in total user watch time. In the competitive landscape of production recommender systems, an improvement of this magnitude is considered a substantial breakthrough.
Conclusion
CREAD demonstrates that the "discretization dilemma" in watch time prediction can be solved through a principled, error-adaptive approach. By balancing learning and restoration errors and enforcing ordinal constraints, the framework provides a robust solution for the extreme imbalances found in real-world video recommendation data.
Summary updated