Scalable analysis and visualization of high-dimensional astronomical data sets

PhD ceremony: mw. B.J. Ferdosi, 13.15 uur, Doopsgezinde Kerk, Oude Boteringestraat 33, Groningen

Dissertation: Scalable analysis and visualization of high-dimensional astronomical data sets

Promotor(s): prof. J.B.T.M. Roerdink, prof. S.C. Trager, prof. J.M. van der Hulst

Faculty: Mathematics and Natural Sciences

In this thesis, Bilis Ferdosi proposes visual and computational paradigms to analyze and extract information from astronomical data. To obtain such techniques, two central issues needed to be addressed: one is the huge size and the other is the large dimensionality of the datasets.

Density estimation approaches can be used for handling large data size. We studied the performance of several density estimation techniques to find a suitable method which is computationally efficient, provides accurate density estimation, and can be used in later stages of data analysis. An interactive approach was developed to find relevant subspaces. The method is strongly tied to morphological properties of object distributions, and is used to identify object clusters. Using this method we recovered various known astronomical relations directly from the data with little or no a priori assumptions.

Using the above method we can identify interesting subspaces of any dimension. However, visualizing high-dimensional structures in a meaningful and user-interpretable way is far from straightforward. For this purpose, we proposed algorithms for reordering dimensions in two widely used high-dimensional data visualization techniques, the parallel coordinate plot and the scatter plot matrix. The effect of reordering is that high-dimensional structures (if present) become easier to perceive. Combining all proposed methods, we designed a visual analytic tool for astronomical data using a large touch-sensitive display.