Decoding a Billion Logs: Quantitative Insights into User Behavior in Internet Video Services

TL;DR

What actually makes a user click "seek" or "close" during a video session? By analyzing over 1 billion logs from a top-tier Chinese video operator, this study quantifies how video length and quality (specifically buffering) dictate user interactions. The verdict: Buffer rate is the king of influence, far outstripping video popularity in determining whether a user stays engaged or starts skipping.

Problem & Motivation: Beyond Traffic Modeling

Historically, researchers looked at Internet video through the lens of bandwidth. We knew how much data was being moved, but we didn't fully understand the why behind user interactions in the wild. Prior works often stayed within controlled environments or small datasets.

The authors of this paper argue that to truly improve Quality of Experience (QoE), we must move beyond intuition. They set out to answer:

1. Does video popularity actually change how we interact with the player?
2. How many minutes do users really spend before dropping off?
3. Which quality metric hurts the user experience the most?

Methodology: The Science of Interaction

The study defines user behavior as a temporal sequence of events ($U_i=\{f_1,f_2,...f_n\}$) occurring within a session. To process the massive dataset, the research team focused on four key attributes:

• User Attributes: Habits and interests.
• Environmental Attributes: Access method and terminal type.
• Content Attributes: Length and popularity.
• Video Quality: Buffer and error rates.

The Core Framework

The researchers employed Information Gain (IG) to determine which factors "explain" user behavior the best. If knowing the buffer rate significantly reduces the uncertainty of whether a user will "seek," then the buffer rate has high Information Gain.

Figure 1: Conceptual model of user interactive behavior over a session timeline.

Key Insights & Experimental Results

1. The "Seek" Dominance

The study found that Seek (Fast-forward/Rewind) is the most frequent interaction, accounting for 77.9% of all user actions. Users are essentially "scanners," constantly hunting for interesting segments.

2. The 20-Minute Retention Rule

Statistical analysis reveals a sharp drop-off in engagement:

• 58.4% of sessions end within 20 minutes.
• Only 1-2% of users remain for longer than an hour.
• Insight: If an operator cannot satisfy a user within the first 20 minutes, they are likely lost.

Figure 2: CDF and Histogram showing the sharp concentration of session lengths in the short-to-medium range.

3. Buffer Rate vs. Popularity

One of the most striking findings is the comparison of influence. Using Information Gain Ratio (IGR), the authors found:

• Buffer Rate influence on "Seek" behavior: 32.6%
• Video Popularity influence on "Seek" behavior: 0.07%

This proves that technical performance (network stability) is orders of magnitude more important for interactive stability than the actual "trendiness" of the content.

4. The Quality-Behavior Correlation

The study used the Least Squares Method to fit curves showing a negative correlation between buffer/error rates and play time. As quality drops (higher buffer rate), the session length collapses predictably.

Figure 3: Correlation analysis showing how increased buffering directly triggers more frequent interactive requests (Seek/Pause).

Critical Analysis & Conclusion

Takeaways for the Industry

• Prioritize Stability: Network operators should focus on reducing buffer events rather than just high bitrates. A smooth 720p stream beats a buffering 4K stream in user retention.
• Interactive Load Prediction: Since seeking increases the demand for network bandwidth (flushing caches), understanding that "Seek" happens most in longer videos allows for better edge-caching strategies.

Limitations

While the dataset is massive, it is limited to a specific geographic region (China) and a specific era of video consumption. The rise of "short-video" (vertical feed) platforms might show different interactive paradigms, such as "swiping" instead of "seeking."

Future Outlook

This work lays the groundwork for automated QoE management. Future systems could theoretically adjustment stream routing or bitrates before a user starts seeking, by observing the early high-frequency interaction patterns that signal user frustration.