BAITRADAR – A clickbait detection deep learning architecture for YouTube

Written by Dr Lim Chern Hong, School of Information Technology

The rising popularity of YouTube led to the clickbait problem, which provokes users to click on videos by publishing attractive titles and thumbnails. It is a waste of time, but the impact is also more severe when it exposes minors to age-inappropriate or ill-intended content. So, it cannot guarantee a safe surfing environment, especially for kids. To detect whether the video is clickbait or not. Hence, this research work uses a Multi-modal Deep Learning architecture as a back-end AI engine to analyse the Title, Tags, Thumbnail, Comments, Audio Transcript and Statistics. We have packaged the model into a browser extension for Chromium-based browsers (Google chrome, edge, brave and others). The browser extension will run the detector in the background and update the user interface (UI) to show the detection result when the user loads the video.

The research significantly impacts the Youtube users (1 billion hours of video watched every day). We aim to improve the clickbait detection problem on Youtube. When the system detects that the video is clickbait, it can alert the user in real-time to avoid wasting their time watching an irrelevant video. Social media companies can also use this to prevent recommending clickbait videos to their customers. Furthermore, since computers can process text-based data more efficiently than video, our proposed method can get a general idea of the whole video in seconds.

The main novelty of our product is the audio transcript to analyse the content of the Youtube video. In addition to using a Multi-modal Deep Learning architecture to combine the Title, Tags, Thumbnail, Comments, Audio Transcript and Statistics for detecting whether a video is clickbait or not. The prototype of the BaitRadar was launched in the Chrome Web Store and is ready to download by users. We published the research at the top conference in signal processing (ICASSP 2021). The product also won a Gold medal in Malaysia's 32nd International Invention, Innovation & Technology Exhibition (ITEX 2021). We will continuously improve the model's performance. We intend to extend the product to support more premium features such as parental control in the future.

Our product's potential markets and stakeholders are general consumers watching YouTube videos and companies/online video platforms planning on using our API to filter YouTube videos with clickbait. The research began with a Final Year Project (FYP) group in the undergraduate program (Bachelor of Computer Science). With Monash University Malaysia's support, technically and financially, we manage to sustain and drive the research to higher achievements. I want to acknowledge my team members Associate Professor Dr Wong Kok Sheik and Mr Bhanuka Manesha Samarasekara Vitharana Gamage, for their significant contributions to the project.