We focus on developing unsupervised clustering techniques applied to the detection of rare facial expressions. Many researchers in human computer interaction, in particular, when setting up computer based eye-tracking studies are primarily interested in identifying the time intervals when the subjects express an emotional attitudinal state. From observing multiple users during simulated studies (which usually take up to 30 minutes) we noticed that only a small fraction of time (perhaps less than 1%) can be labeled as the subject having a clear non-neutral facial expression. This leads us to believe that there is a need for a way to detect these time intervals in an unsupervised manner, which will not require prior training and lengthy initialization.