Structural properties and classification of variable stars: A study through unsupervised machine learning techniques

TL;DR

This study compares PCA and ICA for classifying variable stars using large OGLE datasets, demonstrating ICA's superior performance in identifying resonances and classifying light curves with K-means clustering.

Contribution

The paper introduces an effective combination of ICA and K-means clustering for classifying variable stars and analyzing their light curves, showing improved accuracy over PCA.

Findings

ICA outperforms PCA in resonance detection and classification accuracy.

ICA with K-means effectively constructs period-luminosity and color-magnitude diagrams.

The method is robust across different galaxy datasets (LMC, SMC, MW).

Abstract

The advancement in the field of data science especially in machine learning along with vast databases of variable star projects like the Optical Gravitational Lensing Experiment (OGLE) encourages researchers to analyse as well as classify light curves of different variable stars automatically with efficiency. In the present work, we have demonstrated the relative performances of principal component analysis (PCA) and independent component analysis (ICA) applying to huge databases of OGLE variable star light curves after obtaining 1000 magnitudes between phase 0 to 1 with step length 0.001 for each light curves in identifying resonances for fundamental mode (FU) and first overtone (FO) Cepheids and in the classification of variable stars for Large Magellanic Cloud (LMC), Small Magellanic Cloud (SMC) as well as Milky Way (MW). We have seen that the performance of ICA is better for finding resonances for Cepheid variables as well as for accurately classifying large data sets of light curves than PCA. Using K-means clustering algorithm (CA) with respect to independent components (ICs), we have plotted period-luminosity diagrams and colour-magnitude diagrams separately for LMC, SMC and MW and found that ICA along with K-means CA is a very robust tool for classification as well as future prediction on the nature of light curves of variable stars.