Analysis on gender representation bias of YouTube up-next video recommendation system

Abstract

YouTube is an online video platform on which more than 2 billion monthly users watch around 1 billion hours of video daily. In order to maintain its viewership, YouTube utilizes personalized user-centered sets of recommendation systems. A part of the recommendation system, the 'Up-Next', automatically plays the most relevant video after a video a user has been watching finishes. While the 'Up-Next' function needs to be as equal and fair as the mission of YouTube which is "to give everyone a voice and show them the world", not much research or analysis have been done on the issue. This study proposes a method to measure the level of gender representation in a YouTube video. In total, 7,500 publicly available YouTube videos were collected, from which 618,354 frames were captured for an image analysis. Results show that an average video of 1,200.28 seconds contained 365.37 seconds of male-presenting faces appearing in them whereas only 213.86 seconds of female-presenting faces. While representation of both male- and female-presenting faces declined in long-term, early recommendations made by the YouTube recommendation system navigated users to significantly more male-representative videos.