[Paper Review] Analyzing Network Effects on a Fanfiction Community

Andrés Carvallo, Denis Parra|arXiv (Cornell University)|Sep 6, 2019

Complex Network Analysis Techniques13 references4 citations

TL;DR

This study analyzes fanfiction.net's social network to predict author popularity using logistic regression models that combine user profile data and network metrics. It finds that story count and closeness centrality significantly boost popularity, while time since joining and biography length negatively affect it, with network features improving prediction accuracy by 34% (McFadden R²).

ABSTRACT

Since the early days of the Web 2.0, online communities have been growing quickly and have become important part of life for large number of people. In one of these communities, fanfiction.net, users can read and write stories which are adapted, recreated and modified from original famous books, tv series, movies, among others. By following stories and their authors, the fanfiction community creates a social network. Previous research on online communities has shown how features of the social network can help explain the behavior of the community, so we are interested in studying fanfiction's social network as well as its influence in aspects of the community. In particular, in this article we describe several properties of the members of the community, and we also try to discover which factors explain the popularity of the authors. We discover that time since joining fanfiction and the size of the authors' biography, has a negative effect on the authors popularity. Moreover, we show that the users' network metrics help to explain better authors' popularity.

Motivation & Objective

• To identify factors influencing author popularity in the fanfiction.net online community.
• To investigate whether network metrics such as betweenness, closeness, and PageRank predict author popularity better than profile-based features alone.
• To examine the role of user engagement metrics like time since joining and biography length in determining follower count.
• To compare model performance between profile-only and network-enhanced logistic regression models.
• To understand structural properties of the fanfiction social network, including density and centrality distribution.

Proposed method

• Constructed a social network based on user-favorite author relationships from a sample of 8,877 fanfiction authors.
• Calculated network centrality metrics: betweenness centrality, closeness centrality, PageRank, and clustering coefficient.
• Applied two logistic regression models: one using only profile variables (e.g., stories written, time since joining), and another adding network metrics.
• Used deviance, AIC, and McFadden’s pseudo R² to evaluate model fit and compare predictive power.
• Standardized time and biography word count using natural logarithms to improve model convergence and interpretation.
• Tested significance of coefficients using p-values with conventional thresholds (p < 0.1, 0.05, 0.01).

Experimental results

Research questions

• RQ1 Can network metrics such as closeness centrality and PageRank predict an author’s popularity on fanfiction.net better than profile-based features?
• RQ2 How do profile-level variables—such as number of stories written, time since joining, and biography length—affect an author’s likelihood of having followers?
• RQ3 What is the relative contribution of network structure versus user profile characteristics in explaining author popularity?
• RQ4 Does the inclusion of network metrics significantly improve the predictive power of a logistic regression model for author popularity?
• RQ5 Are there structural patterns in the fanfiction.net network, such as high centrality in a single node or low overall density, that affect popularity dynamics?

Key findings

• The number of stories written is the strongest predictor of author popularity, with a coefficient of 0.88 in the profile-only model and 0.848 in the network-enhanced model.
• Time since joining fanfiction has a significant negative effect on popularity (coefficient = -0.21 in Model 1), suggesting older users are less likely to gain followers.
• Biography length has a negative but non-significant effect on popularity (coefficient = -0.017 in Model 1), indicating longer bios do not increase follower count.
• Closeness centrality has a strong positive effect on popularity (coefficient = 0.994, p < 0.01 in Model 2), making it one of the most influential network metrics.
• Betweenness centrality has a negative and significant effect (coefficient = -0.22, p < 0.01), implying high betweenness is associated with lower popularity.
• The network-enhanced model achieves a McFadden pseudo R² of 0.34, significantly higher than the profile-only model (0.1188), indicating better predictive power when network metrics are included.