[Paper Review] Detection of Fake Users in SMPs Using NLP and Graph Embeddings

Manojit Chakraborty, Shubham Das+1 more|arXiv (Cornell University)|Apr 27, 2021

Spam and Phishing Detection21 references4 citations

TL;DR

This paper proposes a novel spam detection method for Twitter that combines user metadata, NLP-extracted textual features, and graph embeddings from follower-following networks to distinguish genuine from fake users. Using XGBoost on a newly collected, up-to-date dataset of 46,354 active accounts, the approach achieves 96.99% accuracy, outperforming prior work on more recent and challenging data.

ABSTRACT

Social Media Platforms (SMPs) like Facebook, Twitter, Instagram etc. have large user base all around the world that generates huge amount of data every second. This includes a lot of posts by fake and spam users, typically used by many organisations around the globe to have competitive edge over others. In this work, we aim at detecting such user accounts in Twitter using a novel approach. We show how to distinguish between Genuine and Spam accounts in Twitter using a combination of Graph Representation Learning and Natural Language Processing techniques.

Motivation & Objective

• Address the challenge of detecting modern fake and spam accounts on Twitter, which have evolved to evade traditional detection systems.
• Overcome limitations of existing datasets that rely on outdated or deactivated accounts, which no longer reflect current spamming behaviors.
• Develop a robust detection framework using a combination of user metadata, NLP techniques, and graph-based representations to improve detection accuracy on current, active accounts.
• Create a new, up-to-date dataset of 46,354 active Twitter accounts—many of which have passed Twitter’s own spam filters—to train and evaluate the model on realistic, hard-to-detect spam accounts.

Proposed method

• Collected a new dataset of 126,503 Twitter user IDs from public sources and the Twitter API, filtering out inactive accounts to retain 46,354 active accounts with verified labels.
• Extracted 37 user metadata features (e.g., follower count, verified status, account creation date) and full tweet texts for each user via the Twitter API.
• Generated NLP-based features using TF-IDF and BERT-based sentence embeddings to capture linguistic patterns in tweet content.
• Constructed a follower-following graph from user connections and computed graph centrality measures (degree, betweenness, eigenvector, PageRank) and node2vec embeddings to capture structural roles in the network.
• Combined metadata, NLP, and graph features into a unified feature vector (116-dimensional) per user for model input.
• Trained and evaluated multiple supervised ML models, with XGBoost selected as the best-performing model using grid search for hyperparameter tuning (learning rate 0.1, max depth 15).

Experimental results

Research questions

• RQ1 Can a hybrid approach combining NLP, user metadata, and graph embeddings effectively detect modern fake Twitter accounts that have evaded prior detection systems?
• RQ2 How does model performance vary when trained on a new, up-to-date dataset of active accounts—many of which have passed Twitter’s own spam filters—compared to older datasets?
• RQ3 To what extent do graph-based structural features (e.g., centrality, node2vec embeddings) improve spam detection beyond traditional metadata and NLP features?
• RQ4 Does the integration of BERT-based embeddings enhance the model’s ability to detect subtle linguistic patterns in spam content compared to simpler NLP methods?
• RQ5 How does the performance of XGBoost compare to other state-of-the-art models (e.g., SVM, Random Forest, GBM) on this new, challenging dataset of active accounts?

Key findings

• The XGBoost model achieved the highest accuracy of 96.99% on the test set, significantly outperforming other models like Random Forest (94.98%) and LightGBM (88.49%).
• The model achieved 97.90% accuracy on class 1 (spam accounts), indicating strong detection capability for malicious accounts despite their current activity and evasion of platform-level filters.
• The proposed method outperforms prior works that used smaller, older datasets—such as Igawa et al. (96.6%) and Varol et al. (95%)—on a larger and more recent dataset of 46,354 active accounts.
• The model’s F1-score of 0.97 reflects a balanced performance across both genuine and spam classes, indicating high precision and recall.
• The use of node2vec graph embeddings and NLP features (especially BERT-based embeddings) contributed significantly to distinguishing spam patterns, particularly in community-level behavior and linguistic style.
• The dataset’s quality—comprising only active accounts that passed Twitter’s own spam detection—makes the 96.99% accuracy result particularly notable, as it reflects detection of sophisticated, hard-to-detect spam accounts.