Inferring Social Ties: Do Your Twitter Follows Reveal Your Real Friends?

TL;DR

Can the list of celebrities you follow on Twitter reveal who your real-life friends are? This research explores the link between shared interests and social ties. By testing two classical probabilistic models on a dataset of 21,000 users, the authors found that following the same "niche" celebrities (like CEOs or Authors) is a surprisingly accurate predictor of friendship, whereas general celebrity following follows a much more chaotic, less predictable pattern.

Contextual Positioning

In the landscape of social network analysis, this paper serves as a critical bridge between behavioral coincidence and relationship inference. It validates that even "weak" signals—such as following the same public figures—carry enough structural information to de-anonymize or predict "strong" social ties, particularly when those signals are categorized by professional or intellectual interests.

The Core Challenge: Privacy in Public Data

Online behavior is often public, while ties (friendships) are frequently private. Previous work by Kossinets & Watts used email logs, and Crandall et al. used GPS data (Flickr) to infer ties. The challenge here is asymmetry: Twitter is a directed graph where following a celebrity doesn't imply a personal bond, yet the "coincidence" of two people following the same subset of celebrities might suggest they know each other or belong to the same offline community.

Methodology: Two Models, One Question

The authors leveraged a clever temporal split:

1. Ground Truth: Friendships (defined as reciprocal follows) established by July 2009.
2. Activities: Celebrity follows initiated after July 2009 (up to 2013).

This avoids the "circularity" problem where friends follow a celebrity because their friend recommended it, focusing instead on whether shared interest predicts pre-existing social bonds.

The Two Probability Lenses:

• Kossinets & Watts Model: Assumes activities are independent and provides a "baseline" or upper bound. If the actual friendship probability stays below this line, the model is considered a valid bound.
• Crandall et al. Model: A more ambitious model that attempts to provide a "realistic" estimate based on a hypergeometric-style distribution of common items.

Figure 1: Comparison of empirical data against the two models across different categories (Authors, CEOs, etc.).

Deep Dive into Results

The findings offer a nuanced view of social influence:

• The Power of Niche: When users followed specific types of celebrities—CEOs, Authors, or Entrepreneurs—the Crandall model fit the data remarkably well (validated via Kolmogorov-Smirnov tests). This suggests that "professional interest" is a highly structured signal for social clustering.
• The Noise of General Fame: When the model looked at the entire celebrity set (Plot f), the Crandall model failed. Why? Because following "Mega-Celebrities" (Pop stars, movie icons) is a ubiquitous behavior that crosses all social boundaries, making it "noisy" and useless for tie inference.
• Upper Bound Consistency: The Kossinets model held true as a universal upper bound, proving that while friendship is probabilistic, it rarely exceeds the theoretical coincidence limit of independent activities.

Figure 2: Probability of friendship $P(F|C_k)$ as a function of the number of common activities $k$.

Critical Insight & Conclusion

The study proves that homophily (the tendency of individuals to associate with similar others) is quantifiable even through high-level interest data.

Takeaways:

• Categorization Matters: Not all "common activities" are equal. To infer ties accurately, one must weight "exclusive" or "niche" activities more heavily than "popular" ones.
• Privacy Implications: Even if you don't share your friend list, your subscription to technical or niche newsletters/profiles can leak your social circle to an observer.

Limitations:

The researchers noted that the Crandall model, while good for niches, lacked a sufficiently complex "underlying network structure" to handle the chaos of general celebrity following. Future work needs to incorporate social graph topology alongside activity frequency for a more robust prediction engine.