We analyze a portion of the SDSS photometric catalog, consisting of approximately 10,000 objects that have been spectroscopically classified into stars, galaxies, QSOs, late-type stars and unknown objects (spectroscopically unclassified objects, SUOs), in order to investigate the existence and nature of subclasses of the unclassified objects. We use a modified mixture modeling approach that makes use of both labeled and unlabeled data and performs class discovery on the data set. The modeling was done using four colors derived from the SDSS photometry: (u-g), (g-r), (r-i), and (i-z). This technique discovers putative novel classes by identifying compact clusters that largely contain objects from the spectroscopically unclassified class of objects. These clusters are of possible scientific interest because they represent structured groups of outliers, relative to the known object classes. We identify two such well defined subclasses of the SUOs. One subclass contains 58% SUOs, 40% stars, and 2% galaxies, QSOs, and late-type stars. The other contains 91% SUOs, 6% late-type stars, and 3% stars, galaxies, and QSOs. We discuss possible interpretations of these subclasses while also noting some caution must be applied to purely color-based object classifications. As a side benefit of this limited study we also find two distinct classes, consisting largely of galaxies, that coincide with the recently discussed bimodal galaxy color distribution.
تحميل البحث