TL;DR
Researchers have identified a "duration bias" in micro-video recommendation systems where models mistakenly favor longer videos simply because they accumulate more watch time. To fix this, the authors introduce WTG, a new metric that "grades watch time on a curve" based on video length, and DVR, a model-agnostic framework that uses adversarial learning to ensure recommendations are based on genuine interest rather than video duration.
What Problem Does This Paper Address?
In the world of TikTok, Reels, and WeChat Channels, "watch time" is the king of metrics. If a user watches a video for 40 seconds, the algorithm assumes they loved it. However, this creates a massive structural bias: a 60-second video has a much higher "ceiling" for watch time than a 10-second video.
The authors show that current recommender systems suffer from duration bias, leading to two major issues:
- Inaccuracy: Models recommend long videos that users might actually find boring (watching only a small fraction), leading to "bad cases" where a user skips after just a few seconds.
- Unfairness: Creators who specialize in punchy, short content get buried by the algorithm, while long-form "Vlog" style creators receive the lion's share of traffic—not necessarily because their content is better, but because it is longer.
Interestingly, simply switching to "watch percentage" doesn't help; it just flips the bias to favor ultra-short videos.
Methods
The authors propose a two-part solution:
1. Watch Time Gain (WTG)
Instead of looking at raw seconds, WTG standardizes watch time. They divide videos into "bins" based on their duration (e.g., all 15-second videos in one bin, all 60-second videos in another). They then calculate a z-score for a user's watch time relative to the average for that specific bin.
- The Logic: If you watch 10 seconds of a 15-second video, and the average person only watches 5 seconds, your WTG is high. You clearly liked it more than the average viewer of similar-length content.
2. Debiased Video Recommendation (DVR) Framework
DVR is a "wrapper" that can be applied to almost any existing recommendation model (like DeepFM or AutoInt). It employs three strategies:
- Feature Removal: It strips video duration from the input features so the model can't use it as a "shortcut."
- Target Transformation: The model is trained to predict the WTG score rather than raw watch time.
- Adversarial Learning: This is the clever bit. They add a "Duration Regressor" that tries to guess the video's duration based on the model's internal predictions. They then use a Gradient Reversal Layer (GRL) to train the main recommendation model to actively hide duration information from this regressor. This forces the model to learn features representing pure user preference, scrubbed of duration-related noise.
Datasets
The study uses two massive, real-world datasets:
- WeChat Channels: 10,000 users and ~640,000 videos.
- Kuaishou: 20,000 users and ~96,000 videos. Both datasets exhibited strong duration bias in their raw form, with average watch time scaling almost linearly with video length.
Key Insights and Ideas
- Bias Amplification: The authors found that standard machine learning models don't just reflect the bias in the data—they amplify it. The predicted watch time for long videos was often much higher than the actual ground truth, creating a "rich get richer" loop for long content.
- The "Bad Case" Trap: Recommending based on raw watch time leads to a high number of "bad cases" (videos watched for less than 2 seconds). This suggests that the algorithm is frequently "tricked" by long videos that have no actual appeal to the user.
- Model Agnosticism: Because DVR acts as an adversarial layer on top of existing architectures, it doesn't require reinventing the wheel. It improved every backbone model it was tested on.
Key Results
- Massive Performance Gains: When evaluated using the unbiased WTG metric, the DVR framework improved recommendation performance by over 300% compared to standard models in many scenarios.
- Fairness Restored: In the WeChat dataset, standard models gave over 50% of traffic to the longest videos (60s). With WTG/DVR, that traffic was redistributed more equitably, giving short-video creators a fair chance to be seen.
- Reduced Churn Potential: DVR significantly reduced the number of "bad cases," meaning users were less likely to be shown videos they immediately skipped, which is critical for user retention.
Limitations and Future Work
While the paper focuses on watch time, real-world systems are multi-task (optimizing for likes, shares, and follows simultaneously). The authors acknowledge that while these other signals are less prone to duration bias, they are much "sparser" (harder to collect) than watch time.
The authors suggest that the next step is moving from offline simulations to online A/B testing on live platforms to see how these debiased recommendations affect long-term user behavior and creator ecosystems. They also note that while they addressed personalized duration bias, there is still room to explore how aggregate duration bias affects the platform as a whole.
Summary updated