Abstract
Recommender systems are frequently compromised by inherent biases in training data, which prevent models from capturing genuine user interests. In the domain of micro-video recommendation, "duration bias" represents a critical challenge. Because watch time—the primary metric for engagement—is naturally correlated with video length, recommendation models tend to favor longer videos regardless of their actual quality or relevance. This paper identifies the limitations of traditional metrics and introduces Watch Time Gain (WTG), an unbiased evaluation metric that standardizes engagement across varying durations. Furthermore, we propose Debiased Video Recommendation (DVR), a model-agnostic framework that utilizes adversarial learning to decouple user preference from video duration. Extensive experiments on real-world datasets from WeChat Channels and Kuaishou demonstrate that DVR effectively eliminates duration bias and achieves over 30% relative improvement in recommendation performance.
Introduction
In the rapidly evolving landscape of micro-video platforms like TikTok, Reels, and WeChat Channels, "watch time" has emerged as the gold standard for measuring user interest. The logic is intuitive: the longer a user stays on a video, the more they likely enjoyed it. However, this metric contains a structural flaw known as duration bias. A 60-second video has a significantly higher watch-time "ceiling" than a 10-second video. Consequently, recommendation algorithms trained on raw watch time often prioritize longer content, creating a feedback loop that marginalizes short-form creators and leads to suboptimal user experiences.
The Problem: Duration Bias and Its Consequences
Empirical analysis reveals that standard recommender systems do not merely reflect duration bias; they amplify it. This leads to two primary issues:
- Inaccuracy and the "Bad Case" Trap: Models often recommend long videos that users find unappealing. Even if a user skips a long video after a few seconds (a "bad case"), the absolute watch time might still appear comparable to a fully-watched short video, tricking the algorithm into misinterpreting disinterest as engagement.
- Systemic Unfairness: Creators of punchy, high-quality short content are often buried by the algorithm in favor of longer "Vlog" style content, simply because the latter accumulates more seconds by default.
While some systems attempt to use "watch percentage" to counter this, the authors found that this simply flips the bias, unfairly favoring ultra-short videos that are easier to finish.
Proposed Solution: Watch Time Gain (WTG)
To create a level playing field, the researchers introduced Watch Time Gain (WTG). Rather than treating all seconds equally, WTG evaluates watch time relative to the video's duration category.
Videos are grouped into "bins" based on their length. For any given interaction, the system calculates a z-score for the user's watch time relative to the average watch time within that specific bin. This effectively "grades watch time on a curve." If a user watches 10 seconds of a 15-second video when the average viewer only watches 5 seconds, the WTG is high, signaling a strong, genuine preference that is independent of the video's total length.
The DVR Framework
The Debiased Video Recommendation (DVR) framework is a model-agnostic "wrapper" designed to scrub duration bias from existing architectures (such as DeepFM or AutoInt). It employs a three-tiered strategy:
- Feature Removal: Video duration is stripped from the input features to prevent the model from using it as a predictive shortcut.
- Target Transformation: The model is trained to predict the standardized WTG score instead of raw watch time.
- Adversarial Learning: To ensure the model doesn't implicitly learn duration through other features, DVR includes a "Duration Regressor." Using a Gradient Reversal Layer (GRL), the main recommendation model is trained to hide duration-related information from this regressor. This forces the model to learn latent features representing pure user preference, entirely decoupled from video length.
Experimental Results
The framework was tested using two massive real-world datasets: WeChat Channels (10,000 users, 640,000 videos) and Kuaishou (20,000 users, 96,000 videos).
Key Findings:
- Performance Gains: When evaluated using the unbiased WTG metric, the DVR framework improved recommendation performance significantly, showing over 30% relative progress compared to standard models. In specific scenarios, the improvement in precision was even more pronounced.
- Fairness Restoration: In the WeChat dataset, standard models disproportionately allocated over 50% of traffic to the longest videos (60s). DVR redistributed this traffic more equitably, ensuring that short-video creators received a fair share of impressions based on content quality.
- Reduced Churn: By moving away from raw watch time, DVR significantly reduced the frequency of "bad cases" (videos watched for less than 2 seconds), which is a vital metric for maintaining long-term user retention.
Conclusion and Future Work
DVR represents a significant step toward fairer, more accurate micro-video discovery. By standardizing engagement through WTG and enforcing duration-independence via adversarial learning, the framework ensures that user interest, not video length, drives the algorithm.
Future research will focus on integrating these debiasing techniques into multi-task learning environments (optimizing for likes, shares, and follows) and conducting online A/B testing to observe the long-term effects on creator ecosystems and platform-wide user behavior.
Summary updated