Feature extraction using randomwalks

In this paper, a novel idea, which utilizes the metric on a graph, is proposed to extract prominent features for pattern recognition. This proposed model, called “Graphical Metrics Guided Transformation” (GMGT), aims to find projections that can preserve the original metric on the graphic domain in a new Euclidean subspace. With the functional analysis, we present the definition of the metric in the graphical domain and prove that the commute time of random walk is a metric on graphs with the help of real physical model. Furthermore, a new feature extraction algorithm based on GMGT and the commute time is proposed, and is applied to face recognition.