Due to increasing size of astronomical data and expected boom of survey projects, it becomes important to detect interesting objects reliably in the large amount of data. I explain the importance of applying machine learning algorithms in future astronomical research. Focusing on application of clustering algorithms to detect groups in data, I introduce a non-parametric Bayesian clustering method and a consensus clustering method which improve reliability of detecting genuine variable sources in time-series astronomical data. I also present a new strategy of time-series data analysis to identify variable sources quickly by using ensemble of clustering methods as the data size grows. Possible applications of the non-parametric Bayesian method are presented for theoretical and observational astronomical research, emphasising the role of data-driven models.