Abstract:
The rise of clickbait on YouTube undermines user trust and platform credibility by using misleading titles and thumbnails. Current detection methods, which focus mainly on text or engagement metrics, fall short in addressing YouTube’s multimodal content and evolving creator tactics. Additionally, the lack of diverse, multilingual datasets hinders the scalability of existing solutions for a global audience.
This research suggests a real-time, multimodal fusion technique for YouTube clickbait detection. The approach comprises preprocessing numerous data types such as text, thumbnails, and metadata and integrating them into a unified multimodal recognition approach, supplemented by real-time data simulation for performance validation. Therefore, by considering dataset diversity and incorporating post-publishing user engagement metrics, the system ensures adaptation to evolving clickbait techniques and presents scalable and accurate detection capabilities when videos are published.
Preliminary results from the implemented prototype demonstrate the effectiveness of the proposed multimodal fusion approach in detecting Sinhala YouTube clickbait. The fusion model achieved an overall accuracy of 90% for identifying clickbait videos. This result underscore the potential of multimodal fusion in addressing misinformation in low-resource language contexts.
