An improved technique trains a fraud detection system to use mouse movement data as part of a user profile. Along these lines, a training apparatus receives sets of mouse movement datasets generated by a legitimate user and/or a fraudulent user. The training apparatus assigns each mouse movement dataset to a cluster according to one of several combinations of representations, distance metrics, and cluster metrics. By correlating the clusters with the origins of the mouse movement datasets (legitimate or fraudulent user), the training apparatus constructs a robust framework for detecting fraud at least partially based on mouse movement data.