Evaluative Patterns and Incentives in YouTube

Abstract

Users of social media are not only producers and consumers of online content, they also evaluate each other’s content. Some social media include the possibility to down vote or dislike the content posted by other users, posing the risk that users who receive dislikes might be more likely to become inactive, especially if the disliked content is about a person. We analyzed the data on more than 150,000 YouTube videos to understand how video impact and user incentives can be related to the possibility to dislike user content. We processed images related to videos to identify faces and quantify if evaluating content related to people is connected to disliking patterns. We found that videos with faces on their images tend to have less dislikes if they are posted by male users, but the effect is not present for female users. On the contrary, videos with faces and posted by female users attract more views and likes. Analyzing the probability of users to become inactive, we find that receiving dislikes is associated with users becoming inactive. This pattern is stronger when dislikes are given to videos with faces, showing that negative evaluations about people have a stronger association with user inactivity. Our results show that user evaluations in social media are a multi-faceted phenomenon that requires large-scale quantitative analyses, identifying under which conditions users disencourage other users from being active in social media.

Appendix

As a preliminary step to fitting models and testing hypotheses, we survey descriptive statistics to guide the models explained in the previous section. The distributions of views, likes, and dislikes per video are shown on Fig. 3. The histogram of the left panel confirms our observations over the mean and median values of Table 1: all variables are right skewed. This skewness presents heavy right tails that, when eyeballing the plots, suggest the possibility that views, likes, and dislikes follow power-law distributions. Nevertheless, this possibility seems less plausible on the Complementary Cumulative Density Function (CCDF) shown on the right panel of Fig. 3, where the right tails decay faster than it would be expected for a power-law.

Fig. 3.

Video impact distributions. Left: histogram of amount of views, likes, and dislikes over the videos of the dataset. Right: Complementary Cumulative Density Function (CCDF) of counts of videos with more than a certain amount of views, likes, and dislikes. While the histograms show right-skewness, the CCDF of counts show a decay faster than a power-law.

To have a better idea on whether the distributions of amount of views, likes, and dislikes might have scaling properties or diverging moments [25], we applied the method explained in [3] to verify that they do not follow a power-law distribution. We fitted power-law and log-normal distributions to the empirical data, comparing the fits in a log-likelihood ratio test. The results lend very strong evidence favoring the log-normal distribution over the power-law in all three cases: views ($LLR=449.97, p<0.01$), likes ($LLR=275.7, p<0.01$), and dislikes ($LLR=159.99, p<0.01$). This is an example of how informal statistics can be misleading in deciding whether distributions follow a power-law [3, 8], suggesting that we should assume the distributions as log-normally distributed instead.

Fig. 4.

Histograms of log-transformed video metrics. The upper panels (A,B,C) show the histogram of log-transformed amount of views, likes ratio ($L_R$), and dislikes ratio ($D_R$) over the videos of the dataset. Panel D shows the histograms of log-transformed time intervals between videos of the same user ($\Delta t$), in seconds. The vertical red line shows the threshold of inactivity of 2 months. Panel E shows the histogram of time intervals normalized over the average time between videos of the user.

To ensure that we analyze the evaluative tendencies of videos and not their intrinsic correlation with video popularity, we divide likes and dislikes by the amount of views in the variables $L_R=likes/views$ and $D_R=dislikes/views$. These two variables and the amount of views are all roughly log-normally distributed, as it can be appreciated on the histograms of log-transformed values shown on the upper panels of Fig. 4. Some minor skewness can be attributed to integer approximations and boundary values. To cope with these possible deviations from normality in our models, we perform regression diagnostics to model fits to check that residuals are approximately normally distributed.

Figure 4D shows the distribution of the logarithm of the time between videos of the same user $log(\Delta t)$. A clear bimodality is present, but it disappears when normalizing over the average time between videos of each user $\langle t \rangle$. Figure 4E shows the distribution of $log(\Delta t/\langle t \rangle )$, where no bimodality can be observed. This points to the source of bimodality being a variable at the user level, i.e. the activity rate of each user, as the distribution of time intervals collapses to a unimodal distribution after normalization. In our mixed effects regression models of P(I), we include random effects in the form of an intercept for each user that correct for this pattern, ensuring that our results are not a confound with idiographic properties of the users.