Multistage probabilistic classification

We present a probabilistic framework for combining classifiers. Each classifier stage may be based on different data (e.g. astrometry, BP/RP spectrum, G-magnitude, Galactic latitude) or types of object (Galactic, extragalactic). Different data types have differing degrees of power in determining the classes. We present two types of combination. The first uses simple probability theory to combine classifications based on spectroscopy only and astrometry only. The second combines two spectral-only classifiers – one for all classes, the other only for Galactic objects (single and binary stars) – using a weighting function which depends on the astrometry. We introduce a simple approach to simulating Galactic astrometry based on the gamma distribution. We use this to train and test an astrometric-only classifier, one stage in a multistage process.