Exploiting Consensus Clustering for Light Curve Data Analysis

Consensus clustering has been one of the major fields in data science, with increasing numbers of theoretical development and publications over the past twenty years. Recently, a new method for ensemble generation has been introduced with a good use of noise to create diversity via data perturbation. Based on good results with several benchmark data sets, its application to domain-specific problem such as astronomy seems to be an appropriate step ahead. Henceforth, this paper presents an empirical study of the noise-induced consensus clustering with a real data collection, obtained from published LSST light curve catalogue. Note that light curve profiles can be categorized into groups of known astronomical objects with common characteristics and behavior over time. As such, it is important to recognize new or unforeseen objects detected in a sky survey as one of those types, leading to appropriate data collection and further analysis. In particular, two different feature extraction techniques are used to derive features from raw time series records. With these, the performance of simple clustering of k-means and noise-induced ensemble counterpart are compared, using the set of four common clustering validity indices. The experimental results are highlighted with respect to factors of imbalanced data, quality of extracted features and number of clusters. These may help to improve the application of a single or ensemble clustering to light curve data in the future.