Identifying bad guys is critical in applications ranging from fraud detection, to the identification of serial credit defaulters, to law enforcement, to counterterrorism. Machine learning methods can help by profiling bad behavior, so that it can be recognized in the future, and by profiling normal behavior, so that deviations from it can be recognized. Such profiling is challenging for the state of the art, because the data are temporal, multi-relational, networked (people interact), and may be costly to obtain. In this talk, I'll introduce the problems and some techniques for addressing the challenges.