COSMOS continues to advance global scholarship in AI and social network analytics with a new study published in the Journal of Social Network Analysis and Mining, a leading Springer Nature journal recognized for rigorous research at the intersection of network science, data analysis, and societal impact.

In this article, COSMOS researchers introduce SEPS (Semi-Supervised Embedding-based Propagation Scoring), a novel framework for detecting anomalous YouTube channels. By leveraging co-commenter networks, networks of users who comment together across multiple videos, the method combines structural and engagement features to uncover channels exhibiting suspicious behavior, including manipulation of engagement metrics, toxic content dissemination, and potential adversarial information campaigns.

Using a large-scale dataset from the Indo-Pacific region, spanning 97 channels, 702,160 videos, 12.5 million commenters, and 123.9 million comments, SEPS identifies previously unknown anomalous channels while requiring only a small number of labeled examples to guide the learning process. The model employs graph neural networks to generate embeddings for each channel and propagates partial labels through a classification head, effectively separating anomalous from normal channels.

Key findings reveal that SEPS outperforms previous detection methods across recall, precision, and F1-score metrics. Even when only a handful of known anomalous channels are provided, the model maintains high cluster purity, keeping unlabeled anomalous channels grouped together while minimizing false positives. Synthetic dataset experiments further validate the approach, showing its robustness and adaptability to larger, dynamic networks.

The research highlights how anomalous behavior on YouTube is not solely tied to individual content or user activity but emerges from collective interaction patterns. By examining co-commenter networks, SEPS captures subtle signals of coordinated or suspicious activity that may otherwise go unnoticed, offering a powerful tool for platform governance and content integrity.

This work underscores COSMOS’s leadership in AI-powered social media analytics, combining graph-based machine learning, computational social science, and ethical AI to tackle pressing challenges in digital platforms. By publishing in the Journal of Social Network Analysis and Mining, COSMOS continues to shape international discourse on algorithmic transparency, anomaly detection, and the resilience of online ecosystems. Read the full article here.