Online Classification for Time-Domain Astronomy

The advent of synoptic sky surveys has spurred the development of techniques for real-time classification of astronomical sources in order to ensure timely follow-up with appropriate instruments. Previous work has focused on algorithm selection or improved light curve representations, and naively convert light curves into structured feature sets without regard for the time span or phase of the light curves. In this paper, we highlight the violation of a fundamental machine learning assumption that occurs when archival light curves with long observational time spans are used to train classifiers that are applied to light curves with fewer observations. We propose two solutions to deal with the mismatch in the time spans of training and test light curves. The first is the use of classifier committees where each classifier is trained on light curves of different observational time spans. Only the committee member whose training set matches the test light curve time span is invoked for classification. The second solution uses hierarchical classifiers that are able to predict source types both individually and by sub-group, so that the user can trade-off an earlier, more robust classification with classification granularity. We test both methods using light curves from the MACHO survey, and demonstrate their usefulness in improving performance over similar methods that naively train on all available archival data.