Climate data presents unique challenges for machine learning due to its spatiotemporal nature and high-dimensionality. In this talk, I will discuss two applications of high-dimensional modeling for climate data analysis. The first application is on abrupt change detection, with emphasis on detecting significant droughts in the past century. The problem is formalized as a graph-structured linear program (GSLP), and solved using KL-ADM, a novel parallel inexact alternating directions method with Bethe entropy based augmentation. KL-ADM is provably guaranteed to solve GSLPs, and is efficient in practice. When applied to precipitation data over the past century, it detects all major droughts worldwide. The second application is on predictive modeling of land variables based onocean variables. The problem is formalized as a high-dimensional regression problem with hierarchical sparse regularization. Consistency and rates-of-convergence of the estimator will be discussed. When applied to land temperature and precipitation regression problems from 9 different regions worldwide, theproposed methodology is shown to consistently outperform baseline methods.
Thursday, April 19, 2012
Free and open to the public