Seminar on Computational Learning and Adaptation

  A Novel Approach to Novelty Detection

Jeff Scargle
Space Science Division
NASA Ames Research Division

In this talk, I present a brief overview of the problem of detecting novel events, outliers, and anomalies, then discuss the effectiveness of Voronoi tessellations of multiparametric data spaces and their relevance to this problem. After this, I describe a new, simple algorithm for identifying outliers that takes advantages of these structures. I demonstrate results on tasks taken from the literature, including medical diagnosis and ball-bearing fault detection. But the application that inspired this work is the prompt detection of anomalies in water distribution systems, which has applications in maintaining water quality and recognizing bioterrorism events, so I discuss its use there as well.

Return to the seminar schedule