Finding unusual correlation using matrix decompositions

Abstract

One important aspect of terrorism detection is the ability to detect small-scale, local correlations against a background of large-scale, diffuse correlations. Several matrix decompositions transform correlation into other properties: for example, Singular Value Decomposition (SVD) transforms correlation into proximity, and SemiDiscrete Decomposition (SDD) transforms correlation into regions of increased density. Both matrix decompositions are effective at detecting local correlation in this setting, but they are much more effective when combined. © Springer-Verlag Berlin Heidelberg 2004.