[Paper Review] Exposing Paid Opinion Manipulation Trolls

Todor Mihaylov, Ivan Koychev+2 more|arXiv (Cornell University)|Sep 26, 2021

Topic Modeling24 references42 citations

TL;DR

The paper investigates detecting paid opinion manipulation trolls by training on users labeled as trolls by others and testing on paid trolls, using a large feature set derived from user activity to classify trolls vs non-trolls.

ABSTRACT

Recently, Web forums have been invaded by opinion manipulation trolls. Some trolls try to influence the other users driven by their own convictions, while in other cases they can be organized and paid, e.g., by a political party or a PR agency that gives them specific instructions what to write. Finding paid trolls automatically using machine learning is a hard task, as there is no enough training data to train a classifier; yet some test data is possible to obtain, as these trolls are sometimes caught and widely exposed. In this paper, we solve the training data problem by assuming that a user who is called a troll by several different people is likely to be such, and one who has never been called a troll is unlikely to be such. We compare the profiles of (i) paid trolls vs. (ii)"mentioned" trolls vs. (iii) non-trolls, and we further show that a classifier trained to distinguish (ii) from (iii) does quite well also at telling apart (i) from (iii).

Motivation & Objective

• Motivate the need to detect paid trolls in online forums and understand their behavioral patterns.
• Leverage publicly exposed trolls to train a detector despite limited labeled paid-troll data.
• Evaluate how well models trained on 'mentioned' trolls generalize to paid trolls versus non-trolls.

Proposed method

• Construct user-feature vectors from activity history (comments, days active, publications commented on).
• Develop 338 scaled features plus non-scaled equivalents spanning vote-based, similarity, timing, and interaction features.
• Train an SVM with RBF kernel (C=32, gamma=0.0078125) on the 'mentioned' trolls vs non-trolls dataset.
• Test the trained model on four known paid trolls (with 100+ posts) vs four non-trolls to assess generalization to paid trolls.
• Evaluate feature groups by ablation to identify which features drive detection performance.

Experimental results

Research questions

• RQ1 Can a classifier trained on 'mentioned' trolls generalize to identify paid trolls among non-trolls?
• RQ2 Which feature groups most contribute to detecting paid trolls in forum data?
• RQ3 How does model performance vary with the number of posts and the definition of trolls (mentioned vs paid)?

Key findings

• Using all scaled features yields the best overall performance with accuracy 0.88, precision 1.00, recall 0.75, and F-score 0.86 for paid trolls vs non-trolls.
• Time-related and vote-based features are crucial; removing these features degrades precision/recall, with vote features being essential (zero precision/recall if all vote-related features are removed).
• Individual feature groups show that day-of-week, time-related features, and working-hours posting contribute notably to detection, while some groups (e.g., reply-status, similarity) have limited standalone impact.
• Across ablations, testing with paid trolls (100+ posts) is feasible when trained on 'mentioned' trolls, but performance declines for trolls with fewer posts (less than ~40 comments).
• Aggregated profiles indicate paid trolls post less frequently (lower active-days rate) but with concentrated activity during working days/hours, and exhibit higher negative voting on their comments compared to non-trolls.