Few-Shot Learning for Classifying Genuine and Bot Comments on YouTube Using Transformer Models

Abstract

This study aims to develop a comment classification system on the YouTube platform to distinguish between real accounts and bot accounts, addressing the challenge of limited labeled data through a few-shot learning approach. The issue of bot accounts masquerading as real users in comment sections is becoming increasingly prevalent and has the potential to spread spam, misinformation, and influence public opinion. In this study, a Transformer-based model, DistilBERT, is used, which is known for its efficiency in understanding natural language context. The model is trained in a few-shot scenario (N5 to N50) using a very limited amount of training data. Testing results show that the model maintains high and stable performance even with minimal data (N5), achieving an F1-score above 0.90. In addition, this system is implemented into a web application using Flask to enable direct and interactive comment detection. The main contribution of this research is the proof that the combination of few-shot learning and the DistilBERT model can provide a practical and efficient solution for classifying YouTube bot account comments even with limited data conditions, as well as providing a replicable approach for similar problems on other digital platforms.