Fake views removal and popularity on YouTube

Abstract

This paper analyses how YouTube authenticates engagement metrics and, more specifically, how the platform corrects view counts by removing "fake views" (i.e., views considered artificial or illegitimate by the platform). Working with one and a half years of data extracted from a thousand French YouTube channels, we show the massive extent of the corrections done by YouTube, which concern the large majority of the channels and over 78% of the videos in our corpus. Our analysis shows that corrections are not done continuously as videos collect new views, but instead occur in batches, generally around 5 p.m. every day. More significantly, most corrections occur relatively late in the life of the videos, after they have reached most of their audience, and the delay in correction is not independent of the final popularity of videos: videos corrected later in their life are more popular on average than those corrected earlier. We discuss the probable causes of this phenomenon and its possible negative consequences on content diffusion. By inflating view counts, fake views could make videos appear more popular than they are and unwarrantedly encourage their recommendation, thus potentially altering the public debate on the platform. This could have implications on the spread of online misinformation, but their in-depth exploration requires first-hand information on view corrections, which YouTube does not provide through its API. This paper presents a series of experimental techniques to work around this limitation, offering a practical contribution to the study of online attention cycles (as described in the "Data and methods" section). At the same time, this paper is also a call for greater transparency by YouTube and other online platforms about information with crucial implications for the quality of online debate.

Figure 1

(A) 5 sample videos and their hourly evolution of views. (B) Fraction of corrections over real views (histogram) and number of views (blue line) for the 20 most corrected channels in terms of fraction of corrections. (C) Fraction of corrections over real views (histogram) and number of views (blue line) for the 20 most viewed channels in the dataset. (D) Lorenz curve of the distribution of views and corrections among different videos. (E) Percentage of videos affected by corrections.

Figure 2

(A) Distribution of interventions per hour of the day. (B) Distribution of corrections per hour of the day. (C) Distribution of views per hour of the day. (D) Normalized number of interventions, corrections and views per hour following publication. (E) Normalized number of interventions, corrections and views per day after publication.

Figure 3

(A) Fraction of corrections occurring after different percentages of real views. (B) Correlation between real and fake views per channel.

Figure 4

(A) Distribution of interventions in the days following the publication. (B) Distribution of popularity for videos corrected earlier, later and uncorrected.

Figure 5

Error introduced by the benchmark method, varying the time window.

Figure 6

XGBoost parameter tuning. Performances in terms of F1 score associated with different combinations of parameters’ values.

Figure 7

Consistency with 5-min dataset. (A–C) Analysis on the distribution of intervention and corrections. (D,E) Analysis on the relation between speed of correction and popularity.