An Analysis on a YouTube-like UGC site with Enhanced Social Features

Adele Lu Jia, Siqi Shen, Shengling Chen, Dongsheng Li, Alexandru IosupWWW 2017 · DOI

TL;DR

This paper analyzes BiliBili, a UGC platform with rich social features, to understand how these features influence user behavior and content popularity. They introduce novel graph models to capture social structures and demonstrate that these social signals significantly improve the prediction of videos likely to become popular, which is crucial for efficient resource allocation in large-scale UGC systems.

What Problem Does This Paper Address?

Modern User Generated Content (UGC) platforms like YouTube are massive, hosting billions of videos and users. This scale presents a dual challenge: efficiently providing resources (like server capacity for streaming) to ensure a smooth user experience, and designing engaging features to retain and attract users. Traditional approaches to resource provision often overlook the rich social interactions happening on these platforms. This paper aims to bridge this gap by investigating how enhanced social features on a UGC site can be leveraged to predict content popularity and thus inform resource allocation strategies.

Methods

The authors employ a multi-faceted approach:

  • Data Collection and Characterization: They collected a comprehensive dataset from BiliBili, a YouTube-like platform with unique social features (chat replays, virtual donations, non-reciprocal following). They then characterized the platform's scale, video repository, and user activities.
  • Graph Modeling: To capture social structures, they propose two graph models:
  • Follow Graph: A directed graph where nodes are users and edges represent following relationships. This captures explicit social connections.
  • Comment Graph: A graph where nodes are users and edges represent co-commenting on videos under specific conditions (to filter out noise from popular videos). This captures more implicit social interactions.
  • Machine Learning for Prediction: They build machine-learned classifiers (specifically Random Forests) to predict which videos will become popular. They systematically evaluate the impact of different feature groups (video characteristics, uploader attributes, social graph properties, viewing activities) on prediction accuracy.

Datasets

The primary dataset used is a comprehensive view of BiliBili, collected on May 19, 2016. This dataset includes:

  • 2,858,844 videos
  • 28,962,041 users

This dataset is made publicly accessible. The authors contrast their complete dataset with previous studies that often relied on sampled data, highlighting the potential for bias in the latter.

Key Insights and Ideas

  • Social Features Drive Engagement: BiliBili's enhanced social features (chat replays, virtual donations, following) appear to foster greater user engagement, leading to higher viewing rates per video compared to platforms like YouTube, despite BiliBili's smaller overall scale.
  • Uploaders vs. Viewers: Users are more inclined to donate virtual currency to original uploaded content, likely as a gesture of support, even if the content isn't as widely viewed as shared content.
  • Follower-Uploader Reinforcement: There's a positive feedback loop between being a follower and being an uploader. Uploaders with followers tend to upload more, and users with followers are more active in uploading.
  • Implicit Social Connections Matter: The "comment graph," which captures users who co-comment on videos, reveals subtle social relationships that are valuable for prediction, complementing the explicit "follow graph."
  • Social Signals are Powerful Predictors: Social features, particularly those derived from the follow and comment graphs, significantly boost the accuracy of predicting video popularity.

Key Results

  • Dataset Scale: The study utilizes a large-scale, comprehensive dataset of BiliBili, covering over 2.8 million videos and 28 million users.
  • Video Popularity Skew: Video popularity is highly skewed, but BiliBili's distribution is less severe than YouTube's, potentially due to its social features reducing disparity.
  • Classifier Performance:
    • Using only video characteristics yields an AUC of 59.29%.
  • Adding social features dramatically improves prediction accuracy.
  • A classifier using video characteristics, uploader attributes, and follow graph properties achieves an AUC of 84.46%.
  • The best performing classifier, incorporating all feature groups (video, uploader, commentor, follow graph, comment graph, and viewing activities), reaches an AUC of 91.47%. This indicates a high capability to identify videos needing priority in resource provision.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that the "comment graph" construction uses thresholds (t=2, m=50) to filter interactions, and a deeper analysis of these thresholds' sensitivity is left for future work.
  • While the classifier achieves high accuracy, the paper focuses on identifying priority videos rather than proposing specific resource provision methods.
  • Future Work:
    • Further analysis of the enhanced social features and their impact on community engagement.
  • Investigating the sensitivity of the comment graph construction parameters.
  • Exploring specific resource provision methods that can leverage the identified priority videos.

Summary updated

BibTeX
@inproceedings{jia2017analysis,
  author = {Adele Lu Jia and Siqi Shen and Shengling Chen and Dongsheng Li and Alexandru Iosup},
  title = {{An Analysis on a YouTube-like UGC site with Enhanced Social Features}},
  booktitle = {The Web Conference},
  year = {2017},
  doi = {10.1145/3041021.3053901},
}
Figure