Host Galaxy Based Supernova Classification with Machine Learning

Upcoming optical surveys such as the Large Synoptic Survey Telescope will discover supernovae at rates far out-pacing feasible spectroscopic classification. It is therefore critical that we optimize alternate classification methods using all available information. The use of host galaxy data for classification has not been fully developed, and in particular there is an absence of machine learning methods, despite well-known trends between host galaxy properties and supernova types. Using Pan-STARRS1 Medium-Deep Survey (PS1MDS) images, we trained machine learning algorithms to predict supernovae types solely based on contextual information. In particular, we present a random forest classifier using known host galaxy properties, and a convolutional neural network directly using host galaxy images. We demonstrate the value of these algorithms for classifying between five types of supernovae as well as for separating Ia from core collapse supernovae. Future work includes combining our algorithms with photometric classification pipelines to fully optimize classification.