Quantifying Tactical Risk: A Framework for Statistical Classification Using MECH
Additional Document Info
Springer International Publishing Switzerland 2015. This paper presents a statistical classification framework for classification and prediction of asymmetric conflict (AC) locations. Different data normalization and feature reduction methods are coupled with supervised machine learning training algorithms to train classifiers. A set of 77 features derived from the MECH Model (Monitor, Emplacement, and Control in a Halo) were used to train the classifiers. The framework has been implemented and tested on real-world improvised explosive device and direct fire data collected from the conflict in Afghanistan in 2011-2012. Empirical results show that the classifiers achieve high accuracy, with human behavior-related features (visibility and population) exhibiting the most significant statistical differences. Performance of the classifiers is insensitive to the training algorithms. While performance is positively correlated to the training data size as expected, good performance is achieved with a fairly small amount of training data. Experiments based on cross-region training and prediction also show that classifiers are region dependent.