Fig 1.
Statistics of URL usage in the dataset:
(a) proportion of URLs in a particular category, separately for posts (preceded with “P”) and comments (preceded with “C” and dashed), (b) engagement with the posts containing a URL of a particular category in terms of the percentage of posts having at least one comment, the average number of comments for posts having at least one comments, and the average length of the comments in terms of words, (c–d) proportion of URLs in a particular category in posts and comments, over time.
Fig 2.
Distributions of the proportion of URLs posted by a user that are from a particular category, grouped by users whose URLs have a particular political leaning.
Under each distribution, the mean proportion is shown, and a * is shown between two consecutive groups if their distributions differ using the Kolmogorov-Smirnov two-sample test at p<0.001.
Fig 3.
SHapley Additive exPlanations (SHAP), a game theoretical approach for explaining the contribution of each feature to the final output of a ML model.
The Random Forest model predicts how many (log-normalized) science-related URLs a user has posted in our dataset (here, “high” means more URLs were posted), using behavioral features including the categories of the other URLs they shared, as well as the political leaning of those URLs. Top 100 most popular subreddits are used as features, and all others are summed in “other subreddits.”
Fig 4.
Conditional probability of a first-level comment with a URL containing a certain URL category (columns), given it is in response to a post having a scientific, unreliable, right- or left-leaning URL (rows).
Fig 5.
Case study of select subreddits,
( a) the percentage of URLs having a particular domain category and ( b) the percentage of URLs having a particular political leaning. Statistics are shown separately for posts (P) and comments (C).