Random forest for relational classification with application to terrorist profiling

We study the problem of detecting and profiling terrorists using a combination of an ensemble classifier, namely random forest and relational information. Given a database for a set of individuals characterized by both "local" attributes such as age and criminal background, and "relational" information such as communications among a subset of the individuals, with a subset of the individuals labeled as terrorist or normal people, our task is to design a classifier that captures the patterns of terrorists and achieves good accuracy in predicting the labels of the remaining part of the database. In previous work, a hybrid approach was presented that iteratively applies a flat classifier (such as decision trees, fuzzy clustering) augmented with flattened relational attributes for learning and classification. In the current work, our approach is to use random forest as the "flat" classifier in the terrorist detection setting. Random forest is known to have advantage in handling tasks with high dimensionality in input data. This merit of random forest method is very useful for relational learning if the number of "flattened" relational attributes is quite large, which is indeed the case for the terrorist detection task. We report our experiments on a synthetic terrorist database that compare the prediction accuracy of random forest with two other "flat" classifiers, namely, ordinary decision tree and fuzzy clustering. The experimental results show that random forest outperforms both ordinary decision tree and fuzzy clustering.