YouTube as Three Recommendation Networks. The case of The French Media and Political Sphere
Bilel Benbouzid, Emma Gauthier, Alexis Perrier, Pedro RamaciottiThe Recommender System of YouTube is the subject of intense controversy. Traditional media and activists accuse the algorithm of being favorable to contentious content, such as disinformation and hate speech. In contrast, YouTube creators contend that Recommender System focus on a minority of content which is of economic interest to YouTube, to the detriment of alternative videos treating more original and amateur content. This last critic is known as the Adpocalypse hypothesis. In this context, we conduct an empirical research on the consequences of the algorithmic recommendation on YouTube. Our research is centered on different types of YouTube channels, and how they are given visibility through the use of algorithmic recommendations.
In our study, 1400 French YouTube channels representing the French political and media sphere are related among on three types of networks:
1- The social network, corresponding to the network of channels subscribed by, or recommended by the channels themselves;
2- The network of channels which elicit comments from the same audiences, a network of co-commented channels during a 30-day period;
3- The network of recommendations, where channels are linked if a recommendation for a video of one channel was observed when viewing the video of different channel, during a 30-day period.
These three networks reveal the multiple facets of YouTube: the first one encodes the way in which humans recommend each other's channels, the second one shows the audiences shared by the channels, the third one encodes machine-computed similarity between channels. Our observed and collected recommendations are for a non-personalized recommender, or cold start user. We will first present the three channel networks using a network clusterization methods (Stochastic Block Model inference and Louvain Communitarization), showing the different channel classification they induce. In this presentation, we will also compare these three networks according to the diversity of the content that is most central to them. For diversity measurements, we use a manual categorization of the 1,400 channels. We perform a random walk experiment to compute and compare the distributions of the categories. In our case, Perplexity (related to Shannon’s Entropy) is use as a diversity index. We show that the networks of channels produced by humans (networks of human recommendation and co-commenting) produce higher diversities. We show that concentration is produced around channels of traditional media, rather those treating conspiratorial content (for the cold-start component of recommendation). These results help us structure a discussion on the nature of the public space that is being build on YouTube, and the role that disinformation plays in it.