| Intelligence led policing (IPOL) is introduced as a new form of law enforcement that is guided by analysis of data. IPOL has the potential to strengthen the capacity of law enforcement and fraud detection to a considerable degree. Especially the use of machine-learning classifiers plays an important role. However, researchers and politicians are concerned by increasing intrusions on one s privacy due to uncontrolled use of such classifiers. Civil liberties have to be respected. The goal of this project is to provide practical approaches to reliable instance classification. In this way, the probability of civil liberties intrusions is reduced significantly. Research into reliable classifiers has gained increasing interest in the field of machine learning. A reliable classifier is like an expert in law enforcement: she is rarely wrong and sometimes she states I do not know . For binary classification problems this implies that reliable classifiers have high sensitivity (i.e., few false negatives) and high specificity (i.e., few false positives) at the cost of not classifying uncertain instances. The complementary parts sensitivity and specificity have to stay in balance. Public security forces the balance to the side of high sensitivity, whereas civil rights request high specificity. Our research combines theoretical and experimental work. We plan three research directions. First, we will investigate the algorithmic theory of randomness since it provides many insights for thinking about regularities and relations in data. Second, we will extend previous work on Receiver Operator Characteristic (ROC) analysis to handle unclassified instances. The third research direction is to compare different approaches to reliable classification. |
