Analyze the Trend of Post Replies Based on Linear Regression Model-----take Tianyawebsiteas examples

Jingwen Chen, Shumeng Liao, Yuan YinJournal of the Association for Information Systems 2015

Abstract

In recent years, users spend more time on surfing social networking than ever before. How to make information spread rapidly when facing vast amounts of information? Scholars have conducted research on information dissemination in social networking. On the basis of previous research, the authors divide posts into popular posts and ordinary posts and then use a linear regression model to predict replies at specified times. After comparing the difference between two types of posts, the authors conclude that ordinary posts could become popular posts if they maintain a large number of replies within the first five hours and increase replies by making use of community mechanisms. This conclusion provides a reasonable proposal for enterprises and administrators to identify and recommend popular posts.

Introduction

In the era of Web 2.0, information overload is a constant challenge for both users and platform administrators. For businesses, identifying which posts will become "popular" is essential for brand management and marketing. For administrators, it is vital for content moderation and highlighting high-quality discussions.

The authors identify a gap in existing research: while many models exist for information dissemination, they often fail to account for the specific community mechanisms of sites like Tianya. Furthermore, there is a lack of comparative study between "popular" and "ordinary" posts. This research sets out to determine if early-stage data (the first few hours of replies) can accurately predict the total number of replies a post will have at the end of its 48-hour "life cycle."

Methodology

The researchers employed a log-transformed linear regression model to tackle the non-linear nature of social media engagement.

  1. Classification: They applied the 20/80 rule (Pareto Principle), defining the top 20% of posts by reply count as "popular" and the remaining 80% as "ordinary."
  2. Modeling: Recognizing that the raw relationship between early replies and late replies is not strictly linear, they performed a logarithmic transformation on the data. This allowed for the use of Ordinary Least Squares (OLS) estimation to find a significant linear relationship in the log-log space.
  3. Restoration: To make the results actionable, a specific mathematical correction (incorporating the variance of residuals) was used to transform the logarithmic predictions back into a standard numerical scale.
  4. Tooling: Data was processed using SPSS, and predictive power was tested across various "early" time points (from 1 to 47 hours) against the 48-hour total.

Data Collection

The study used a custom dataset scraped from the Tianya website, specifically the "International Observation" section.

  • Collection Method: A Java-based web crawler captured data every 15 minutes.
  • Sample Size: 645 posts collected between October 23 and October 30, 2014.
  • Scope: The authors focused on the 48-hour window following a post's publication, as observations showed that hits and replies stagnate significantly after this period.

Analysis and Insights

The Long-Tail Distribution

The data confirmed that forum engagement follows a "long-tail" curve. A vast majority of posts receive fewer than 50 replies, while a tiny fraction (the "popular" posts) see counts ranging from 200 to 600. This distribution validates the use of the 20/80 rule for post classification.

Figure 1: Histogram showing the long-tail distribution of post replies at the 48-hour mark.

The 5-Hour Critical Window

For ordinary posts, the "maximum predicted value" for total replies peaks if the post performs well within the first five hours. After this point, if engagement doesn't hit a certain threshold, the predicted total drops sharply. This highlights a critical window for intervention or promotion.

Platform Feedback Loops

The authors attribute the "richer get richer" phenomenon to Tianya’s sorting algorithm. Posts with new replies are bumped to the "default column" (the front page), increasing exposure. This creates a virtuous cycle for popular posts, but means ordinary posts quickly disappear from view if they don't gain immediate traction.

Key Results

  • Threshold for Popularity: In the dataset of 645 posts, the threshold for a "popular" post was 32 replies within 48 hours.
  • Growth Patterns: Popular posts showed a steep, almost linear growth in predicted totals when analyzed at the 18-hour mark, eventually stabilizing around 1,400 replies for the most viral content.
  • Mean Performance: For popular posts, the mean predicted value peaked at the first hour and then declined to a stable average of about 75 replies. For ordinary posts, the mean stabilized at a much lower 6 replies after the 10-hour mark.
Figure 5: Mean predicted value of total replies relative to the predicted time point.
  • Strategic Intervention: The results suggest that if a post can maintain a high volume of replies within the first five hours, it has a significantly higher probability of being pushed into the "popular" category by the site's internal mechanisms.

Conclusion and Future Work

While the log-transformed linear regression model provides strong predictive capabilities, the authors acknowledge several areas for improvement. The current model relies solely on the number of replies and does not account for "hits" (views), the reputation of the original poster, or the sentiment of the content.

Future research should compare these findings with other platforms, such as video-based sites like Youku, to see if the 5-hour window and 20/80 rule hold true across different media types. Additionally, incorporating natural language processing (NLP) to analyze content could reveal if specific keywords or topics inherently drive higher engagement coefficients.

Summary updated

Abstract

In recent years, users spend more time on surfing the social networking than ever before. How to make the information spread rapidlywhen facing vast amounts of information?Scholars have conducted information dissemination in social networking. On the basis of previous research, the authors divide posts into popular posts and ordinary posts and then use the linear regression model to predict the replies at specified time. After comparing the difference between two types of posts, the authors concludethat ordinary posts could become popular posts if the posts could maintain a large number of replies within former five hours and increase replies by making use of community mechanism. This conclusion provides a reasonable proposal for enterprises and administrators to identify andrecommendpopular posts.
BibTeX
@article{chen2015analyze,
  author = {Jingwen Chen and Shumeng Liao and Yuan Yin},
  title = {{Analyze the Trend of Post Replies Based on Linear Regression Model-----take Tianyawebsiteas examples}},
  journal = {Journal of the Association for Information Systems},
  year = {2015},
}
Figure