💻 computer science

Playback Quality Anomalies Associated with Rapid Swiping in Short Videos: Model-Attributed Association Estimation and Risk Prediction

This study leverages a large-scale dataset and a quality-aware machine learning model to accurately distinguish playback-quality issues from low user interest in rapid video swipes, achieving high predictive performance and enabling targeted optimization of streaming experiences.

Original authors: Yuewei Yuan, Tianyi Xu, Jialei Huang, Jiayang Yin

Published 2026-09-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuewei Yuan, Tianyi Xu, Jialei Huang, Jiayang Yin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When you scroll through a short video on your phone, the app is constantly trying to guess what you want to see next. It relies heavily on how long you watch a clip and how quickly you swipe past it. If you watch for a long time, the system assumes you liked it; if you swipe away almost immediately, it assumes you were bored. This logic seems sound, but it misses a crucial detail: sometimes people swipe away instantly not because they dislike the content, but because the video simply refused to play. A slow start, a frozen screen, or a stuttering connection can make a viewer abandon a video they might have otherwise enjoyed. For the algorithms that power these apps, distinguishing between a user who is uninterested and a user who is frustrated by technical glitches is a difficult puzzle. If the system cannot tell the difference, it might stop showing great videos to people simply because their internet connection was slow that day, or it might fail to fix the technical problems that are driving people away.

A team of researchers set out to solve this specific problem by looking at the hidden signals that happen before a user swipes away. They analyzed a massive amount of data from twenty-one days of short-video usage, covering hundreds of millions of viewing sessions. Their goal was to build a system that could look at the technical health of a video playback—such as how long it took for the first image to appear, how many times the video paused to buffer, and the speed of the network connection—and determine if a quick swipe was actually a complaint about quality rather than a rejection of the content. They found that a significant portion of these rapid exits were indeed caused by technical failures. By teaching a computer model to recognize these patterns, they created a way to predict when a user is about to leave due to poor performance, separating these technical failures from genuine lack of interest.

The researchers started by gathering a vast collection of records that linked what users watched with how their phones and networks were performing at that exact moment. They looked at specific moments when a video failed to load smoothly, such as when the first frame took longer than 1.2 seconds to appear, or when the video had to pause multiple times to download more data. They also tracked the device's memory usage and the speed of the connection. Using this information, they trained a model to identify the difference between a user who is just bored and a user who is dealing with a broken playback experience. The model learned to spot a specific set of conditions: a slow start, repeated buffering, a sudden drop in internet speed, or a device running out of memory. When these technical issues happened together, the model could predict with high accuracy that a rapid swipe was likely a reaction to the poor quality, not the video itself.

One of the most important discoveries was how often these technical issues actually drive people away. The study found that under certain difficult conditions, such as when a phone has very little free memory and the internet connection is slow, nearly 35 percent of the rapid swipes were linked to playback problems. This is a substantial number, suggesting that a large chunk of the data the apps use to decide what to show next is actually noise caused by technical glitches. The researchers also found that the time it took for the video to start was a major factor; if the first image took more than 1.2 seconds to appear, the risk of the user swiping away increased significantly. Similarly, if the video had to buffer two or more times in a row, the user was likely to leave, and this frustration often carried over to the next video they tried to watch.

To make sure their model was reliable, the researchers tested it against a set of data that had been carefully reviewed by human experts. The model proved to be very accurate, correctly identifying the cause of the swipe in most cases. It was also able to assign a confidence score to its predictions, meaning it could tell the system how sure it was that a swipe was due to quality issues. This is important because it allows the system to act only when it is confident, rather than guessing. The model showed that it could distinguish between a user who is genuinely uninterested and one who is frustrated by a slow connection, even when the user's past behavior suggested they liked similar content. This ability to separate the signal from the noise helps the system understand the true preferences of its users.

The researchers then used their findings to simulate what would happen if the apps changed how they delivered videos to users who were at risk of a bad experience. In this simulation, when the model predicted a high risk of a quality-related swipe, the system would try to fix the problem before it happened. It would do this by loading a bit more of the video in advance, lowering the video quality slightly to ensure it played smoothly, or switching to a different server that was closer to the user. The results of this simulation were promising. By making these adjustments, the time it took for the first frame to appear dropped by nearly 22 percent, and the time spent waiting for the video to buffer fell by more than 30 percent. The system also became much better at predicting whether a user would finish watching a video, with the error in those predictions dropping by 37 percent.

This work highlights a fundamental shift in how we understand user behavior on digital platforms. It suggests that not every quick swipe is a sign of boredom; many are simply a reaction to a broken experience. By recognizing the difference, platforms can stop misinterpreting technical failures as a lack of interest. This allows them to fix the underlying problems that cause frustration, ensuring that users see the content they actually want to watch. While the study was conducted using simulations and historical data rather than live changes to a real app, the results provide a clear path forward. The next step would be to test these ideas in the real world to see if fixing these technical glitches leads to happier users and better recommendations. For now, the research offers a powerful tool for understanding the invisible friction that keeps people from enjoying the content they seek.

Technical Summary: Playback Quality Anomalies Associated with Rapid Swiping in Short Videos

Problem Statement
Short-video recommender systems traditionally interpret rapid swiping (exits within seconds) as a direct signal of low user interest or content preference. However, this behavioral signal is often confounded by playback quality issues, such as first-frame delays, buffering interruptions, and network instability. Existing literature has largely addressed watch-time debiasing, noisy behavior correction, or system-level Quality of Experience (QoE) in isolation. There remains a critical gap in distinguishing between preference-driven rapid swiping and swiping caused by playback anomalies at the exposure level. Without this distinction, recommendation signals are misinterpreted, and playback optimization strategies lack targeted precision.

Methodology
The authors propose a unified framework that integrates weakly supervised labeling, model-attributed association estimation, and confidence calibration to estimate the risk of rapid swiping driven by playback quality. The approach is built upon four core modules:

Data Construction and Weak Supervision:
The study utilizes 21 days of de-identified logs comprising 234 million exposures and 58.6 million sessions. Rapid swipes are defined as exits occurring within 1.5 seconds, within 3 seconds, or following a playback interruption. Weakly supervised probabilistic labels are generated based on these exit behaviors combined with eight specific anomaly thresholds (e.g., first-frame delay >1,200 ms, ≥2 buffering events, cumulative buffering >1,000 ms, RTT increase >50%, bandwidth decrease >40%, memory usage >85%, error codes, and failure to resume within 1 second). These labels serve as soft targets for training, with boundary cases (0.40–0.60 confidence) retained for robustness.
Model-Attributed Association Estimation:
A dual-branch architecture encodes playback quality events (first-frame delay, buffer count, RTT, bandwidth, memory, etc.), short-term viewing behavior, and control variables (content category, creator, recommendation position, user preference). The model employs a channel-masking technique to estimate the contribution of specific quality channels to the rapid-swipe risk. The contribution of a quality channel $j$ for exposure $i$ is calculated as the logit difference between the full model and the model with channel $j$ masked, holding all other inputs fixed. This allows for exposure-level conditional importance estimation without assuming causal effects.
Temporal Risk Prediction:
The core prediction engine utilizes a quality-aware time-series model. It projects input features into a 96-dimensional latent space and processes them through a three-level dilated causal convolutional encoder (with dilation rates of 1, 2, and 4) followed by an 8-head attention mechanism. This structure captures cross-exposure dependencies and temporal anomalies (e.g., successive buffering events). The model outputs a rapid-swipe probability, which is further refined using a joint loss function that balances binary cross-entropy, KL divergence (for weak labels), Brier score, and ordinal consistency among different exit types.
Confidence Calibration:
To address probability bias and overconfidence, a conditional temperature calibrator is fitted on the validation set using temperature scaling. This module adjusts the logits to minimize negative log-likelihood, specifically targeting the Expected Calibration Error (ECE). The calibrated model produces risk scores that more accurately reflect the actual frequency of rapid swiping.

Key Results
The proposed framework was evaluated on an independently adjudicated test subset (12,000 exposures) and compared against baselines including Logistic Regression, XGBoost, LightGBM, Cox models, LSTM, and Transformer architectures.

Performance Metrics: The proposed model achieved an F1 score of 0.831 and an AUC of 0.879 on the adjudicated test set. For rapid-swipe prediction specifically, it achieved an AUC of 0.867 and an AUPRC of 0.704. These results outperformed all baselines, with the F1 score exceeding the best baseline (Transformer-based session model) by 0.014 and the AUC by 0.015.
Attribution Analysis: The model identified that first-frame delays exceeding 1.2 seconds and repeated buffering events are strongly associated with rapid swiping. Under combined low-memory and high-RTT conditions, quality-associated cases accounted for 34.8% of rapid swipes. The attribution ranking was robust, showing high Spearman correlation (0.952) with Integrated Gradients and leave-one-channel-out ablation.
Calibration: The calibration process significantly improved reliability, reducing the Expected Calibration Error (ECE) from 0.146 to 0.052 (a 64.4% relative decrease), while maintaining AUC and AUPRC scores.
Offline Policy Evaluation: In a counterfactual replay simulation, a risk-aware policy (adjusting preloading, bitrate, and CDN routing based on predicted risk scores) reduced the P95 first-frame time by 21.8% (from 1,420 ms to 1,110 ms) and P99 buffering duration by 30.7%.

Significance and Claims
The paper claims that distinguishing quality-related swipes from preference feedback is essential for improving the interpretation of recommendation signals and enabling targeted playback optimization. The primary contribution is a method that extends rapid-swipe modeling from simple preference prediction to quality-aware behavioral attribution. By controlling for content, creator, and user preference, the framework isolates the impact of technical anomalies on user behavior.

The authors emphasize that their work provides interpretable probabilistic evidence for correcting recommendation feedback and optimizing resource allocation (e.g., preloading and bitrate selection). However, the paper maintains a modest stance regarding its conclusions:

The attribution results are described as "robust but not causal."
The offline policy evaluation is explicitly stated as a replay-based estimate conditional on fixed network traces and unchanged user interest, not an online intervention effect.
The authors acknowledge limitations including log-sampling bias, label noise, and restricted scenario coverage, noting that causal effects and cross-platform transferability require further quasi-experimental and online validation.

In summary, the study offers a technical pathway to deconfound user behavior from system performance, suggesting that a significant portion of "disinterest" signals in short-video platforms are actually artifacts of playback quality, which can be mitigated through risk-aware optimization strategies.

Technical Summary: Playback Quality Anomalies Associated with Rapid Swiping in Short Videos

More like this