Real-Time Classification of Transient Events in Synoptic Sky Surveys

An automated rapid classification of the transient events detected in modern synoptic sky surveys is essential for their scientific utility and effective follow-up when resources are scarce. This problem will grow by orders of magnitude with the next generation of surveys. We are exploring a variety of novel automated classification techniques, mostly Bayesian, to respond to those challenges, using the ongoing CRTS sky survey as a testbed. We describe briefly some of the methods used. The increasing number of synoptic surveys is now generating tens to hundreds of transient events per night, and the rates will keep growing, possibly reaching millions of transients per night within a decade or so. Generally, follow-up observations are needed in order to exploit scientifically these data streams to the full. In optical surveys, for instance, all transients look the same when discovered—a starlike object that has changed its brightness significantly—and yet between them they could represent vastly different physical phenomena. Which ones are worthy of a follow-up? This is a critical issue for the massive event streams such as LSST, SKA, etc., and the sheer volume demands an automated approach (Donalek et al. 2008; Mahabal et al. 2010; Djorgovski et al. 2011a). The process of scientific measurement and discovery operates typically on time-scales from days to decades after the original measurements, feeding back to a new theoretical understanding. However, that clearly will not work when changes occur on time-scales that are shorter than those needed to set up a new round of measurements. It demands real-time systems incorporating a computational analysis and decision engine, and optimized follow-up instruments that can be rapidly deployed with immediate analysis and feedback, and implies automated classification and decision-making systems. The classification process for a given transient involves: (1) obtaining available contextual archival information, and combining it with the measured parameters from the discovery pipeline, (2) determining (relative?) probabilities or likelihoods of it belonging to some class of transient, (3) obtaining follow-up observations to disambiguate competing classes, (4) using those as a feedback and repeating for an improved classification. We describe below a few techniques that help in this process. Our principal dataset is the transient event stream from the Catalina Real-time Transient Survey (CRTS; http://crts.caltech.edu; Drake et al. 1999; Djorgovski et al. 2011b; Mahabal et al. 2011), but the methodology we are developing is more universally applicable. Bayesian Networks. The available data for any given event would generally be heterogeneous and incomplete. That is difficult to accommodate in the standard machine-