Deciphering the Pulse of Video Popularity: An Epidemic Lens on Recommendation Mechanisms

TL;DR

Understanding how a video goes viral—whether through a "push" from the platform or a "pull" from social sharing—is notoriously difficult. This paper introduces a novel framework using epidemic models to mine video view count traces. By splitting influence into Direct Recommendation and Word-of-Mouth (WOM), the authors provide a mathematical microscope to evaluate why certain videos succeed while others stall.

Background: Beyond the Black Box

Why do we care? For platforms like Netflix or YouTube, knowing the ROI of a front-page feature versus the organic growth of a viral clip is the billion-dollar question. Traditional metrics are often implicit or segmented. This study treats view count evolution as a physical process of "infection," where information flows through a population of potential viewers.

Problem & Motivation: The Resource Constraint

Existing popularity models often treat information diffusion as an infinite process. However, in the real world, recommendation resources are finite. A homepage only has so many slots. The authors identified that prior works (like the EvoModel) failed to capture this "resource-aware" nature, leading to poor fitting of real-world traces where a video's growth abruptly plateaus or starts with a significant delay ($t_s$).

Methodology: The CModel Framework

The core innovation is the Combined Model (CModel), which treats the total cumulative views $x(t)$ as the sum of two distinct epidemic flows:

1. Direct Model (DModel): Represents deliberate promotion by the provider. It is characterized by a "push" rate $\alpha$ and a starting time $t_s$.
2. Word-of-Mouth Model (WModel): Represents the "viral" component where one viewer "infects" others with a social sharing coefficient $\beta$.

The authors mathematically prove that without sufficient "seed" users or a high enough social sharing rate ($\beta$), organic growth will likely fail, necessitating the "direct" intervention of the platform.

The Equation (8) demonstrates the non-linear ODE used to capture the social diffusion characteristics.

Experiments & Results: Real-World Validation

The models were tested on a massive dataset from Tencent Video, covering Movies, TV series, News, and Music Videos (MV).

SOTA Comparison

Compared to 7 other benchmarks (Ad-hoc, Weibull, etc.), CModel showed significantly higher robustness. For instance, in Movie categories, CModel had 0% failed cases (NMSE > 10), whereas Ad-hoc models failed in over 58% of scenarios.

Key Insights from Mining

• The Power of the Push: For nearly all categories, Direct Recommendation ($N_D$) outweighed WOM ($N_W$). Platforms are the primary drivers of traffic.
• Time Sensitivity: News videos have the shortest $t_s$ (delay), meaning they are pushed immediately. Movies often have a longer delay while the system gathers early user feedback.
• The WOM Threshold: The authors show a "concave/convex" relationship for $\beta$, proving that unless social sharing hits a specific threshold, it is highly inefficient for diffusing information.

Figure 4. Scatter plots showing the correlation between Direct and WOM views across different content types.

Critical Analysis & Conclusion

Takeaway

The CModel provides a "reverse engineering" tool for video platforms. By fitting historical data to this model, a manager can identify "under-recommended" videos—those with high intrinsic attractiveness ($q$) and high viral coefficients ($\beta$) that were never given enough direct push ($\alpha$).

Limitations

The model struggles with extremely unpopular videos (under 1000 views) because their traces are too noisy. It also assumes a fixed total potential population ($N$), which might not hold for long-tail videos that stay relevant for years.

Future Outlook

This work paves the way for "active" recommendation. Instead of just reacting to views, systems could use the early-stage fitting of $\beta$ and $q$ to predict which videos deserve a "premium" spot on the front page, optimizing the allocation of finite recommendation resources.