Fig. 4: Controversy and toxicity in conversations.
From: Persistent interaction patterns across social media platforms and over time
a, The mean controversy (σ(l)) and mean toxicity versus thread size (log-binned and normalized) for the Facebook news, Twitter news, Twitter vaccines and Gab feed datasets. Here toxicity is calculated in the same conversations in which controversy could be computed (Extended Data Table 3); the relative Pearson’s, Spearman’s and Kendall’s correlation coefficients are also provided in Extended Data Table 3. Trends are reported with their 95% confidence interval. b, Likes/upvotes versus toxicity (linearly binned). c, An example (Voat politics dataset) of the distributions of the frequency of toxic comments in threads before (n = 2,201, minimum = 0, maximum = 1, lower whisker = 0, Q1 = 0, Q2 = 0.15, Q3 = 0.313, upper whisker = 0.769) at the peak (n = 2,798, minimum = 0, maximum = 0.8, lower whisker = 0, Q1 = 0.125, Q2 = 0.196, Q3 = 0.282, upper whisker = 0.513) and after the peak (n = 2,791, minimum = 0, maximum = 1, lower whisker = 0, Q1 = 0.129, Q2 = 0.200, Q3 = 0.282, upper whisker = 0.500) of activity, as detected by Kleinberg’s burst detection algorithm.
