Introduction
As the "Social Web" expanded in the late 2000s, the volume of socially generated metadata—tags, reviews, and comments—grew exponentially. Traditionally, researchers assessed the "quality" of this content using external experts. However, this approach fails to capture how a specific community actually values its own content. This paper investigates how the Digg community self-regulates by analyzing what makes a user comment "good" or "bad" according to the community's own voting system.
Methods and Dataset
The researchers utilized a custom dataset crawled in November 2008, consisting of 4,500 stories containing 232,000 comments from 38,000 unique users. They targeted older stories to ensure voting activity had stabilized.
To understand community preference, three categories of features were analyzed:
- Visibility: The "Digg score" of the parent article and the chronological order of the comment.
- User Reputation: Metrics such as the number of articles a user submitted and their history of received ratings.
- Content Analysis: Linguistic analysis measuring length, complexity (Entropy), readability (SMOG score), and sentiment (subjectivity and polarity).
The study framed the problem as a classification task, sorting comments into four buckets: Excellent (>600 score), Good (1–600), Fair (-100–0), and Bad (<-100).
Key Insights
The analysis revealed several surprising dynamics within the Digg ecosystem:
- The Power User Paradox: While a small group of "power users" often controlled which news stories reached the front page, user reputation proved to be a poor predictor of comment success. Being a "celebrity" on the platform did not guarantee high ratings for individual comments.
- The "First-Mover" Advantage: Visibility is paramount. Comments posted earlier in a thread's lifecycle received significantly higher mean scores than those buried at the end.
- Preference for Simplicity: The community favored brevity and clarity. Comments with lower linguistic complexity and higher readability scores tended to perform better.
Results and Predictive Modeling
To predict comment outcomes, the researchers tested Linear Regression and a Quadratic Classifier. The Quadratic Classifier achieved an overall accuracy of 85%. It was particularly effective at distinguishing between "Fair" and "Good" comments (precision of 0.82 and 0.70, respectively).
However, predicting "Excellent" comments proved difficult (precision of 0.02). This suggests that while "good" comments follow predictable patterns of simplicity and timing, top-tier comments likely possess a "viral" or "witty" quality that standard linguistic features cannot capture.
Conclusion
This research highlights the potential for platform self-regulation. By understanding the internal mechanics of community preference, social platforms can build filters that promote high-quality content based on community values rather than arbitrary external standards. Future work may involve incorporating more nuanced features to capture the "spark" of high-scoring comments and exploring the relationship between comment and article sentiment.
Summary updated