Kathryn Linehan

CUR Matrix Approximation Using Convex Optimization

Monday, March 25, at 1:00pm
 
Abstract: 
In this talk we present a CUR matrix approximation that uses a novel convex optimization formulation to select the columns and rows of the data matrix for inclusion in C and R, respectively.  We discuss implementation of the algorithm using the surrogate functional of Daubechies et al. [Communications on Pure and Applied Mathematics, 57.11 (2004)] and extend the theoretical guarantees of this approach to our formulation.  Applications using CUR as a feature selection method for classification will be shown, if time.  In addition, the proximal operator of the L-infinity norm is used in our CUR algorithm.  We present a neural network approximation to this proximal operator that uses a novel feature selection process based on moments of the input data in order to allow vectors of varying lengths to be input into the network.