This talk examines two separate statistical problems for which low-dimensional models are effective.
In the first part of the talk, we examine the robust PCA problem: given a matrix M that is the sum of a low-rank matrix L and a sparse noise matrix S, recover L and S.
This problem appears in various settings, including image processing, computer vision, and graphical models. Various polynomial-time heuristics and algorithms have been proposed to solve this problem under certain conditions. We introduce a block coordinate descent algorithm for this problem and prove a convergence result. In addition, our iterative algorithm has low complexity per iteration and empirically performs well on synthetic datasets.
In the second part of the talk, we examine a variant of ridge regression: unlike in the classical setting where we know that the parameter of interest lies near a single point, we instead only know that it lies near a known low-dimensional subspace.
We formulate this regression problem as a convex optimization problem, and introduce an efficient block coordinate descent algorithm for solving it.
We demonstrate that this "subspace prior" version of ridge regression is an appropriate model for understanding player effectiveness in basketball. In particular, we apply our algorithm to real-world data and demonstrate empirically that it produces a more accurate model of player effectiveness by showing that (1) the algorithm outperforms existing approaches and (2) it leads to a profitable betting strategy.