Detecting anomalies in graph networks on digital markets

Area of Science:

• Computer Science
• Data Science
• Network Security

Background:

• Digital markets face threats from anomalous activities, including bots and fraudulent users.
• Existing methods for anomaly detection often lack computational efficiency and scalability.
• Information asymmetry and inadequate market actor protection are persistent challenges.

Purpose of the Study:

• To evaluate and compare various graph-based methods for detecting anomalies in digital markets.
• To propose an efficient and robust algorithm for threat detection that improves market actor protection.
• To reduce information asymmetry through enhanced anomaly detection.

Main Methods:

• Utilized three publicly accessible graph-based datasets: TwiBot-20 (Twitter social network), Bitcoin OTC, and Bitcoin Alpha (cryptocurrency markets).
• Compared different graph algorithms, focusing on node statistics versus state-of-the-art graph embeddings for feature extraction.
• Developed and proposed a new graph-based statistics algorithm for anomaly detection.

Main Results:

• Graph-based data significantly outperforms text data in predicting anomalies.
• Node statistics-based methods demonstrate superior predictive performance and computational efficiency compared to graph embeddings.
• The proposed algorithm based on node statistics offers a more efficient and less overhead-intensive solution than graph embeddings, which are dataset-specific and require extensive tuning.

Conclusions:

• Graph-based statistics methods are highly effective for anomaly detection in digital markets, offering better performance and efficiency than graph embeddings.
• The newly proposed algorithm provides a scalable, robust, and computationally efficient solution for threat detection.
• The findings advocate for node statistics-based approaches to enhance digital market security and reduce information asymmetry.