Abstract: In this talk, I will discuss a convex formulation of the clustering problem and its generalization to biclustering of matrices and more broadly to co-clustering of multiway arrays or tensor data. The key advantage in formulating clustering as a convex program is that doing so addresses well-known issues of instability and parameter selection that plague mainstream approaches. The formulation also admits a simple first order iterative algorithm for solving the problem with global convergence guarantees. We also provide a finite sample bound for the prediction error of our convex co-clustering (CoCo) estimator. Finally, we illustrate the utility of this formulation and algorithm in biclustering high throughput bioinformatics data and in co-clustering click-through contingency tables of online advertising data.

Short Bio: I am an assistant professor in the Department of Statistics at North Carolina State University. Before coming to NC State, I earned my PhD in statistics at Rice University in 2011. After completing my PhD, I worked as a postdoctoral researcher in both the Human Genetics department at UCLA and the Digital Signal Processing group at Rice University. My research interests are in statistical learning and numerical optimization and their application to analyzing large and complicated modern data in biological science and engineering applications.