ETD

• anomaly detection
• bot detection
• semi-supervised learning
• supervised learning
• Telegram
• time series analysis
• unsupervised learning

This thesis investigates the problem of inauthentic engagement on Telegram, where bots inflate post view counts and mislead advertisers who pay for visibility that no real audience sees. The growing reliance on engagement metrics to assess content popularity makes this type of manipulation particularly damaging, as inflated statistics are difficult to detect without looking at how views accumulate over time. The work proposes a pipeline that combines unsupervised clustering and anomaly detection with supervised classification to identify posts with manipulated view counts, using the temporal shape of view growth as the main signal. Experiments were performed on a dataset collected from public Telegram channels, and results suggest that temporal patterns alone carry enough signal to distinguish suspicious posts from genuine ones.