An investigation into label fusion on sparse data

The Department of Homeland Security (DHS) and the Transportation Security Administration (TSA) are interested in developing a standardized testing procedure for determining the performance of a candidate detection systems combined with a currently operational detection system. This paper investigates the problem of combining data from different detection systems that could possibly produce a more accurate combined detection system. Theoretically, this combination of detection systems tends to assume that the Receiver Operating Characteristic (ROC) curves that describe the detection systems are well defined for all thresholds on the unit square. In practice, many ROC curves are discrete and/or incomplete data sets and label fusion with a competing existing system is not well defined. This work investigates the concept of fusion with incomplete ROC data using an estimate for the fused performance of the detection system under a constrained OR label rule.