Comparison of Stellar Classification Accuracies Using Automated Algorithms

Machine learning algorithms can be used to reduce the dimensionality of stellar spectra, and make automated classification using these spectra faster and more efficient. Much work has been done in applying one such algorithm, Principal Component Analysis (PCA), to spectra, but PCA assumes that the data is linear globally, which is not the case for spectral data. Other work has been done to examine the possibilities of applying nonlinear dimension reduction algorithms, such as Locally Linear Embedding (LLE) and Isometric Mapping (Isomap). These algorithms only require local linearity in the data, which is satisfied by spectral data. We apply PCA, LLE and Isomap to spectra of K-type stars from SDSS, and use a trained Support Vector Machine to determine subclasses of these stars. We then compare these subclasses to those given by SDSS to determine the accuracy of the classifications. This allows us to compare the classification accuracies of PCA, LLE, and Isomap.