Co-commenters as clues: a partial-label approach to detecting anomalous channels on Youtube Article Swipe
YOU?
·
· 2025
· Open Access
·
· DOI: https://doi.org/10.1007/s13278-025-01493-0
· OA: W4412753896
We propose a semi-supervised approach that finds anomalous YouTube channels by combining co-commenter networks and engagement features. Our method, named SEPS (Semi-Supervised Embedding-based Propagation Scoring), uses a small set of labeled anomalies to guide the learning process. It trains a graph neural network to embed nodes, spreads those partial labels through a simple classification head, and flags channels that appear suspicious. The model was tested on real data from the Indo-Pacific region which spans 97 channels, 702,160 videos, 12.5 million commenters, and 123.9 million comments. Further tests on synthetic data show that SEPS outperforms older methods in recall, precision, and F1-score. It detects more unlabeled anomalies with fewer false alarms. We also show that it remains effective when only a handful of known anomalies are present, and it can maintain high cluster purity under these conditions. In addition, we discuss how the model can scale to larger datasets and adapt to changes in user behavior. This work contributes a robust way to spot suspicious channels in real-world environments where full labels are scarce.
