The good, the bad and the bait: detecting and characterizing clickbait on youtube

Abstract

The use of deceptive techniques in user-generated video portals is ubiquitous. Unscrupulous uploaders deliberately mislabel video descriptors aiming at increasing their views and subsequently their ad revenue. This problem, usually referred to as 'clickbait,' may severely undermine user experience. In this work, we study the clickbait problem on YouTube by collecting metadata for 206k videos. To address it, we devise a deep learning model based on variational autoencoders that supports the diverse modalities of data that videos include. The proposed model relies on a limited amount of manually labeled data to classify a large corpus of unlabeled data. Our evaluation indicates that the proposed model offers improved performance when compared to other conventional models. Our analysis of the collected data indicates that YouTube recommendation engine does not take into account clickbait. Thus, it is susceptible to recommending misleading videos to users.