Hybrid AI models for variable star classification

Abstract

The classification of variable stars is an important problem for many people in the astronomy community. Variable stars provide a unique means to probe the internal structure of distant stars. The process of classifying these stars presents itself perfectly to machine learning. This paper proposes a hybrid model, combining the feature extraction capabilities of a convolutional neural network with the classification power of a non-parametric kernel density estimation paired with a naive-Bayes classifier. On classification tasks with 15 types of variable stars, the model achieves roughly 77% accuracy, placing it in line with the average performance seen throughout the literature. The hybrid model is special, though, in that it achieves this accuracy with comparatively less training time, a smaller sample size, and more specific classes.