Knowledge Discovery in Mega-Spectra Archives

The recent progress of astronomical instrumentation resulted in the construction of multi-object spectrographs with hundreds to thousands of micro-slits or optical fibres allowing the acquisition of tens of thousands of spectra of celestial objects per observing night. Currently there are two spectroscopic surveys containing millions of spectra. These surveys are being processed by automatic pipelines, spectrum by spectrum, in order to estimate physical parameters of individual objects resulting in extensive catalogues, used typically to construct the better models of space-kinematic structure and evolution of the Universe or its subsystems. Such surveys are, however, very good source of homogenised, pre-processed data for application of machine learning techniques common in Astroinformatics. We present challenges of knowledge discovery in such surveys as well as practical examples of machine learning based on specific shapes of spectral features used in searching for new candidates of interesting astronomical objects, namely Be and B [e] stars and quasars.