Decentralized view prediction for global content placement

Stéphane Delbruel, Davide Frey, François Taïani

Abstract: This paper proposes a decentralized architecture for predicting the regional popularity of user-generated content using video tags and peer-to-peer (P2P) overlays. By leveraging self-organizing networks to sample local "tastes," the system enables proactive content placement across different countries without a central authority, improving scalability and reducing server load.

Introduction

Modern video traffic is dominated by massive, centralized Content Delivery Networks (CDNs). While effective, these systems favor large-scale providers and require immense infrastructure. Decentralized alternatives—where users’ own devices, such as set-top boxes, provide storage—offer a more scalable and democratic solution. However, these systems face a significant "global knowledge" problem: to place a video where it will be most viewed, a system typically requires a global database of viewing patterns. In a fully decentralized network, no single node possesses this comprehensive view.

This research, originally presented at the 2015 Middleware ARM Workshop, addresses this challenge by enabling global predictions using only local information. The proposed system allows a decentralized network to self-organize content placement efficiently by predicting a video's future popularity across different regions before it goes viral.

Decentralized Architecture

The core of the approach is a tag-based prediction model adapted for a decentralized environment. The system relies on three primary components:

  1. Tag-based Fitness: The model assumes that if videos with a specific tag (e.g., "soccer") are popular in a region (e.g., Brazil), a new video with the same tag will likely be popular there as well. A "fitness" score is calculated for a video in a specific country based on the average views of its tags in that region.
  2. Self-Organizing Overlays: Each country maintains its own P2P overlay network. Within these overlays, users are connected to others with similar tastes, determined by a similarity metric of their viewing history.
  3. Virtual Nodes and Probing: When a user uploads a video, the system uses a Random Peer Sampling (RPS) protocol to find an "entry point" node in various target countries. This node acts as a "virtual node" that probes its local country-level overlay to estimate potential interest.
Decentralized network architecture showing cross-country probing and content placement.

The Trend Graph Model

Because querying every user in a country is impossible for scalability, the system employs a sampling heuristic. The virtual node identifies its $k$ most similar neighbors within the regional overlay. By plotting the similarity scores of these top neighbors, the system creates a "Trend Graph."

The Trend Graph uses linear regression on the top-k most similar nodes to estimate regional interest.

As shown in the figure above, the system applies linear regression to these $k$ best nodes. The area under this regression line serves as a mathematical proxy for the total predicted view count (fitness) of the entire country. This allows the network to approximate global interest using only a small, local sample of the population.

Key Insights and Results

  • User-as-a-Video: A significant conceptual contribution is treating a user’s profile—the sum of all tags they have watched—as a "paradigmatic video." This allows the same mathematical similarity metrics to be used for comparing users to videos and users to each other.
  • Proactive Placement: Unlike most P2P systems that move content reactively after it becomes popular (caching), this system moves content at the time of upload (placement). This proactive approach aims to eliminate initial lag and "flash crowd" bottlenecks.
  • Mathematical Framework: The primary result of this architectural study is the derivation of a method to rewrite global fitness equations into a summation of individual tastes that can be estimated via P2P sampling.

Limitations and Future Work

While the framework is promising, several challenges remain. The current model uses a simple linear approximation for the Trend Graph; higher-order approximations may be required for increased accuracy. Additionally, the system is highly dependent on the quality of video tags; poorly tagged content or rare tags may lead to prediction failures.

Future work involves integrating temporal dynamics, such as "sliding windows," to favor recent trends over older data. The authors are also focused on extensive experimental validation using real-world YouTube traces to compare these decentralized heuristics against full-information centralized models.

Summary updated

BibTeX
@inproceedings{delbruel2015decentralized,
  author = {Stéphane Delbruel and Davide Frey and François Taïani},
  title = {{Decentralized view prediction for global content placement}},
  booktitle = {Arm@middleware},
  year = {2015},
  doi = {10.1145/2834965.2834974},
}
Figure