Time-Series Classification for Astronomical Surveys

Start: 02/24/2011 - 4:15pm
End  : 02/24/2011 - 5:00pm

Statistics/OR/Math Finance Seminar

Joseph Richards, UC Berkely


Next-generation astronomical surveys require sophisticated classification tools for optimal allocation of follow-up resources and to construct pure and complete samples of objects for cosmological inference.  These surveys retrieve light curves for millions of astronomical sources, such as variable stars, quasars, and supernovae.  Each light curve is a time series of unevenly-spaced, often sparse, measurements of an object's brightness.  The challenge of light-curve classification, then, is two-fold: first, sets of class-predictive features must be estimated from each light curve, and second, flexible classification models that can handle both high-dimensional feature spaces and the veritable zoo of astronomical objects need to be employed.  I will detail my use of diffusion maps and Lomb-Scargle periodograms for light-curve feature estimation, and random forests and structured classification models for multi-class light-curve classification.  I describe my work on two classification problems: supernova typing using simulations from the Dark Energy Survey Supernova Classification Challenge and variable star classification using light curves from a few existing surveys.  I will conclude by presenting some thoughts on methods to remedy sample-selection bias in astronomical classification problems.

Harvey Mudd College 3rd floor Sprague. Refreshments at 4pm.