Modeling terrorism culpability: An event-based approach

Abstract

Recently, researchers have become interested in the issue of assessing culpability for terrorist attacks when no one group claims or multiple groups claim responsibility. Several new methods have been put forward for predicting culpability, traditionally assessed by intelligence analysts, using both machine learning and statistical classification models. These models have had varying degrees of success, with new ensemble classification models performing generally better than traditional statistical techniques. This paper applies a relatively new methodology, Random Forests, to the problem of predicting culpability and compares it to some of the more frequently used statistical classification techniques, including multinomial logistic regression and naïve Bayesian classification. Though generally outperforming other techniques, Random Forests struggles with unbalanced data, performing worse than either of the other models tested in the class with the least information. However, this evaluation of Random Forests for the assessment of terrorism culpability is positive. Implications of the model and comparison to other models are discussed and ways forward are suggested.