Abstract
Predicting the popularity of online videos is a critical yet difficult task for content platforms, primarily due to sudden viewership spikes driven by recommendation algorithms. This paper introduces a signal-processing approach using the Discrete Fourier Transform (DFT) to classify video view traces. By analyzing high-frequency energy, the authors separate predictable "smooth" videos from erratic "fluctuating" ones. The study demonstrates that for fluctuating videos, a counterintuitive strategy of discarding historical data prior to a major popularity peak significantly improves prediction performance across multiple standard models.
Introduction
Content providers such as YouTube and Tencent Video rely on accurate popularity predictions for backend resource provisioning, advertising pricing, and recommendation logic. Most existing predictive models assume that a video's future growth is a smooth continuation of its past. However, modern platforms often feature videos on homepages or in trending sections, causing massive, sudden spikes in views. These "sharp turns" in viewership traces break traditional models. This research aims to quantify predictability, identify why certain videos defy standard forecasting, and propose a method to handle erratic viewership patterns.
Methodology
The core of the proposed approach is the application of frequency domain analysis to social metrics.
Frequency Domain Analysis
The authors apply a Discrete Fourier Transform (DFT) to 365-day daily view count traces. By converting time-series data into the frequency domain, they can isolate high-frequency components that correspond to sharp jumps and drops in viewership.
Predictability Classification
A predictability metric, $\tau$, is calculated as the ratio of high-frequency energy to total energy. Videos are then classified into two groups:
- Group_S (Smooth): Videos with a low $\tau$ ratio, indicating predictable growth.
- Group_F (Fluctuating): Videos exceeding a specific energy threshold, indicating hard-to-predict, erratic behavior.
Refined Prediction Strategy
For the erratic videos in Group_F, the authors propose a "data pruning" strategy. They identify the day of the maximum view count peak ($t_h$) and discard all data prior to it. This approach assumes that a major spike represents a regime change (e.g., a platform recommendation) that makes previous history irrelevant or even harmful to future predictions.
Experimental Results
The methodology was validated using a proprietary Tencent Video dataset (8,955 videos) and a public Sina Weibo dataset (0.3 million messages).
Model Validation
The classification split was tested against five standard prediction models: Szabo-Huberman, Multivariate Linear, MRBF, RPP-ex, and ARIMA. In all cases, videos in Group_S yielded significantly lower Mean Bounded Relative Error (MBRE) than those in Group_F. By applying the data-discarding strategy to the RPP-ex model for Group_F, prediction errors dropped dramatically, bringing the performance on erratic videos close to the baseline performance on smooth videos.
Comparison with Microblogs
Analysis of Sina Weibo data revealed that microblog repost traces are generally "smooth" compared to video views. Approximately 66% of microblogs fell into Group_S. The authors hypothesize that microblog virality is more organic and distributed (driven by user retweets), whereas video virality is often centralized and artificial (driven by platform-wide recommendations).
Discussion and Conclusion
The study highlights that "more data" is not always better; historical data from a different viewership regime can actively mislead predictive models. Using FFT as a lightweight heuristic allows platforms to route different types of viewership traces to optimized modeling pipelines. While the heuristic thresholds for frequency boundaries require careful selection, the approach provides a robust framework for improving the reliability of resource provisioning and recommendation systems in the face of algorithmic volatility.
Summary updated