GROUSE is an incremental algorithm for subspace identification based on incomplete information, proposed and studied by Laura Balzano, Rob Nowak, and Ben Recht at Madison. This talk discusses recent results on the local convergence behavior of GROUSE, showing an expected linear convergence rate. Stronger results are possible when the full data vector is available. We note too an equivalence between GROUSE and an incremental method based on the singular value decomposition.
Jeremy Weiss’s abstract:
Accurate prediction of future onset of disease from Electronic Health Records has important clinical and economic implications. In this domain, we are presented with timelines, i.e., the arrival of events comes at semi-irregular intervals, which makes the prediction task challenging. Continuous-time Bayesian networks effectively model such processes but are limited by the number of conditional intensity matrices, which grows exponentially in the number of parents per variable. We discuss our work developing a partition-based representation and using regression trees and forests whose parameter sizes grow linearly in the number of node splits. Using a multiplicative assumption we show how to update the forest likelihood in closed form, producing efficient model updates. Multiplicative forests can be learned from few timelines with gains in performance and scalability particularly when risk factors are known to be independent, or multiplicative in nature.
Discovery Building, Orchard View Room
Jeremy Weiss, Stephen Wright