[Paper Review] Followers or Phantoms? An Anatomy of Purchased Twitter Followers

Anupama Aggarwal, Ponnurangam Kumaraguru|arXiv (Cornell University)|Aug 7, 2014

Spam and Phishing Detection19 references16 citations

TL;DR

This study analyzes purchased Twitter followers from underground markets, identifying behavioral and profile traits that distinguish them from legitimate users. Using supervised learning with behavioral features—especially unfollow entropy—it achieves 88.2% accuracy in detecting suspicious follow behavior, revealing key indicators of fake follower networks.

ABSTRACT

Online Social Media (OSM) is extensively used by contemporary Internet users to communicate, socialize and disseminate information. This has led to the creation of a distinct online social identity which in turn has created the need of online social reputation management techniques. A significant percentage of OSM users utilize various methods to drive and manage their reputation on OSM. This has given rise to underground markets which buy/sell fraudulent accounts, `likes', `comments' (Facebook, Instagram) and `followers' (Twitter) to artificially boost their social reputation. In this study, we present an anatomy of purchased followers on Twitter and their behaviour. We illustrate in detail the profile characteristics, content sharing and behavioural patterns of purchased follower accounts. Previous studies have analyzed the purchased follower markets and customers. Ours is the first study which analyzes the anatomy of purchased followers accounts. Some of the key insights of our study show that purchased followers have a very high unfollow entropy rate and low social engagement with their friends. In addition, we noticed that purchased follower accounts have significant difference in their interaction and content sharing patterns in comparison to random Twitter users. We also found that underground markets do not follow their service policies and guarantees they provide to customer. Our study highlights the key identifiers for suspicious follow behaviour. We then built a supervised learning mechanism to predict suspicious follower behaviour with 88.2% accuracy. We believe that understanding the anatomy and characteristics of purchased followers can help detect suspicious follower behaviour and fraudulent accounts to a larger extent.

Motivation & Objective

• To understand the behavioral, profile, and content-sharing characteristics of purchased Twitter followers from underground markets.
• To identify discriminative features that distinguish fake followers from legitimate users.
• To develop a supervised learning model capable of detecting suspicious following behavior with high accuracy.
• To evaluate the reliability of market claims about follower quality and retention.
• To provide actionable indicators for detecting fraudulent follower networks on Twitter.

Proposed method

• Collected purchased follower accounts from an underground Twitter follower market using a no-followback scheme to avoid credential compromise.
• Extracted profile attributes (e.g., bio, profile image, language) and behavioral features (e.g., unfollow frequency, RT and @mention engagement rates).
• Defined 'unfollow entropy' as a metric to quantify irregular unfollowing behavior over time.
• Used a Support Vector Machine (SVM) with RBF kernel for classification, trained on a 70-30 split with 10-fold cross-validation.
• Incrementally added feature sets (profile, behavioral, content-based) to evaluate their individual and collective impact on detection accuracy.
• Evaluated model performance using standard metrics: accuracy, F1-score, and AUC, with confusion matrix analysis.

Experimental results

Research questions

• RQ1 What are the distinguishing profile and behavioral characteristics of purchased Twitter followers compared to legitimate users?
• RQ2 How do purchased followers differ in content sharing and interaction patterns from random Twitter users?
• RQ3 To what extent do underground markets uphold their service guarantees regarding follower quality and retention?
• RQ4 Which behavioral features are most effective in identifying suspicious following behavior?
• RQ5 Can a supervised learning model accurately classify suspicious follower behavior using measurable user attributes?

Key findings

• Purchased followers exhibit a high unfollow entropy rate, indicating frequent and irregular unfollowing of followed accounts, a behavior uncommon among legitimate users.
• These accounts show significantly lower social engagement with their friends, including low RT and @mention engagement ratios.
• The study found that underground markets do not consistently deliver on their promises of high-quality, active, or long-retained followers.
• The supervised learning model achieved 88.2% accuracy in detecting suspicious follower behavior, with the highest performance gain from behavioral features.
• Unfollow entropy, RT-engagement ratio, @mention-engagement ratio, language overlap, and social reputation were among the most informative features for detection.
• Profile-based features (e.g., bio, profile image) were the least effective for identifying suspicious behavior, likely due to sparse or absent user-generated content in legitimate accounts.