 |
Center for Nonlinear Studies |
For numerous problems involving prediction and classification from data, a productive approach has been to consider a graph representation of the data, and solve the problem using network classification methods. I will discuss two instances of this. The first is an ongoing collaboration with LANL researchers studying fracture systems in subsurface rock. By considering topological features of an underlying graph that describes the fracture network, we use supervised learning methods to provide rapid and accurate predictions of regions of high particle transport. The second problem concerns automated classification of vehicles from roadside audio sensors. For this problem, we form a graph where nodes represent frequency-domain descriptions of time windows of the recorded audio signal. I will illustrate how a straightforward application of spectral clustering reduces dimensionality to the point where k-nearest neighbors provides almost perfect vehicle class identification.
Host: Gowri Srinivasan