From playing games to committing crimes: A multi-technique approach to predicting key actors on an online gaming forum

Abstract

We propose a systematic framework for analysing forum datasets, which contain minimal structure, and are nontrivial to analyse at scale, aiming to support future analysis of underground forum communities. We use a multi-technique approach which draws on a combination of features, including post classifications extracted using natural language processing tools, and apply clustering and predictive techniques to this dataset, to predict potential key actors-individuals who have a central role in overtly criminal activities, or activities which could lead to later offending, and hence might benefit most from interventions. We predict 49 key actors on an underground gaming-specific cheating and hacking forum, validated by observing only overlaps of techniques, combined with topic analysis, to build a classifier for key actor status. In addition, we also use these techniques to provide further insight of key actor activity. We found one cluster and two posting trajectories to contain a high proportion of key actors, logistic regression found an actor's h-index to have higher odds for prediction than other features, and partial dependence plots found reputation to have a significant change in prediction between values of 100 to 1000.