Behavioral Tracing of Twitter Accounts

Neel GuhaPrivOn@ISWC 2017 · OpenReview

Abstract

Online platforms struggle with a persistent cycle of malicious behavior where banned users simply create new pseudonymous accounts. This paper introduces "behavioral tracing," a novel method for account linking that identifies whether two distinct Twitter accounts belong to the same individual by analyzing the semantic signatures of their posts. By utilizing Doc2Vec for vector representation and a custom weighted similarity function, the system can link accounts with high precision without relying on invasive metadata. Our results show that incorporating retweet history and penalizing generic topical interests significantly improves identification accuracy, achieving an F1 score of 0.69.

Introduction

Online communities face a "perpetual cycle" of disruption. When a malicious actor or "troll" is banned, the low barrier to entry allows them to register a new account and resume their activities immediately. Current defense mechanisms, such as human validation or technical blacklisting (IP/phone verification), are often resource-intensive, easily bypassed, or potentially harmful to the privacy of legitimate users in sensitive contexts.

The challenge lies in "account linking"—determining if a new account is a returning banned user based solely on public activity. While prior research has focused on network graphs or specific lexical clues, this work addresses the gap by focusing on semantic signatures: the underlying topical interests and behavioral traces that remain consistent even when a user changes their handle or social circle.

Methodology

The behavioral tracing pipeline consists of three primary stages:

  1. Vector Representation (Doc2Vec): Rather than analyzing individual tweets, all posts from an account are collated into a single document. Using Doc2Vec, these documents are mapped into a high-dimensional vector space where accounts with similar topical interests are positioned in close proximity.
  2. Neighborhood Discovery (K-Means): To ensure scalability, K-means clustering is used to group similar accounts. Detailed comparisons are only performed between accounts within the same "neighborhood," reducing the computational overhead of exhaustive pairwise matching.
  3. Weighted Similarity Scoring: Standard cosine similarity often fails because many users share "generic" interests (e.g., news, weather). The core innovation is a weighted similarity function that rewards accounts for shared niche interests while penalizing similarity to the "bulk" of the dataset. This effectively filters out common noise to isolate unique individual signatures.

Experimental Setup and Dataset

The study utilized a dataset collected via the Twitter API consisting of 1,270,999 tweets from 1,849 users. To ensure the model relied on semantics rather than network metadata, all account handles (e.g., "@user") were stripped from the text. For evaluation, each user's tweet history was split into two separate "accounts" to test the algorithm's ability to re-link them.

Cumulative distribution of total tweets per account in the dataset.

As shown in the tweet distribution, the dataset contains a wide range of activity levels. Analysis suggests that the algorithm's performance stabilizes at a threshold of roughly 200 tweets, after which it hits a point of diminishing returns.

Key Results

The behavioral tracing method significantly outperformed both random baselines and standard clustering techniques. The inclusion of retweets proved to be a critical factor; the F1 score jumped from 0.54 to 0.69 when retweets were included. This suggests that a user's retweet history serves as a highly specific "composition of semantic signatures" representing their unique information diet.

Performance comparison of behavioral tracing against baselines.

The weighted similarity approach provided a nearly 50% improvement over raw K-means clustering on the same Doc2Vec vectors, highlighting the importance of distinguishing unique behavioral traits from generic user behavior.

Discussion and Limitations

Behavioral tracing treats a person's interests and information processing style as an immutable identifier, making it a potent tool for persistent identification. However, several challenges remain:

  • Temporal Drift: Human interests evolve over time, potentially making older traces less effective for linking to current activity.
  • Context Switching: Users maintaining accounts for entirely different purposes (e.g., professional vs. personal) may not exhibit enough semantic overlap to be linked.
  • Scalability: As the dataset grows, the likelihood of overlapping fingerprints in the vector space increases, which may slightly degrade the algorithm's ability to extract a unique signature.
  • Adversarial Adaptation: Sophisticated users might eventually attempt to "poison" their own semantic trace to avoid detection, leading to a potential cat-and-mouse game in future security applications.

Summary updated

BibTeX
@inproceedings{guha2017behavioral,
  author = {Neel Guha},
  title = {{Behavioral Tracing of Twitter Accounts}},
  booktitle = {PrivOn@ISWC},
  year = {2017},
  url = {https://openreview.net/forum?id=ZB5w70EaD9},
}
Figure