Identification of Deceptive Clickbait Youtube Videos Using Multimodal Features

Abstract

In social media platforms like Youtube, clickbait content has become a standardized method of driving public attention to video content for the creator’s benefit. With the gradual standardization of clickbait on Youtube, an increasing trend is observed for the probability of deceptive intentions within the content. Smarter clickbait and masquerading techniques allow fake sensationalized media information to generate public attention. Thus, it is harder for users on video platforms to distinguish between misleading and legitimate videos. This work explores the deception aspect of clickbait on Youtube. It proposes a clickbait classification technique exploiting various DL and ML models on the multimodal features of the collected clickbait videos. This work also developed a Youtube video corpus containing recent clickbait videos and devised techniques for utilizing multimodal features such as video text titles, video thumbnails, and video metadata properties using natural language processing techniques. We have explored standard deep learning techniques such as CNN, LSTM, and Bi-LSTM for textual features and ResNet50, VGG16, and VGG19 for visual feature extraction. In the multimodal approach, the combined model CNN $\oplus$ ResNet50 achieved the highest F1-score (74.8%) among all multimodal techniques.