Oral Presentation
Galaxy Group Finding via Unsupervised Clustering
Presenter: Hai-Xia Ma (Nagoya University)
Galaxy groups are collections of galaxies comprising about a few to a hundred gravitationally bounded members. They are potent indicators of the structures of mass at the largest scales, enabling us to assess the features of the Universe. However, galaxy groups usually tend to have chaotic structures. This makes it difficult to determine if galaxies in their region of the sky are gravitationally connected to the group. We propose the use of unsupervised machine learning algorithms to make an unbiased identification of groups.
In this work, we made use of six clustering algorithms, which are K-means, Friends-of-friends, Gaussian Mixture Models, Agglomerative Hierarchical Clustering, Ordering Points to Identify the Clustering Structure (OPTICS), Hierarchical Density-Based Spatial Clustering (HDBSCAN), to test their performance on finding galaxy groups from the galaxy catalogs (De Lucia & Blaizot, 2007) built from the Millennium Simulation (Springel et al, 2005). We defined recovery and complete rates of group finders to optimize the values of hyper-parameters of each algorithm, and evaluate how close they have predicted for the group members comparing to the real halo members in the mock catalog. Of these six algorithms, OPTICS most consistently balances purity rate completeness, even with a high fraction of identified groups which are exactly the same as given in simulated catalog. We conclude that OPTICS is a robust group finder that is effective at determining a wide range of group shapes and sizes with minimal contamination.
asiaa.sinica.edu.tw