On the application of machine learning approaches in astronomy: Exploring novel representations of high-dimensional and complex astronomical data

The goal of the presented work is the application of data-driven methods on complex and high- dimensional astronomical databases. The focus of the work is the exploration of novel data representations in order to enable the use of statistical learning approaches in the analysis of data. With the help of diverse science cases, the advantages of the introduced approaches for classication, visualization and regression tasks are shown by applying the developed methodology to astronomical data. In the first part, an alternative approach for estimating redshifts of spectra by using the knowledge about the redshifts provided by the SDSS pipeline is presented. A novel data repre- sentation is employed which contains only information relevant for estimating the redshift and the detection of multiple redshift systems. Subsequently, a novel data representation for regu- larly sampled light curves based on recurrent networks is presented. This allows an explorative investigation of huge databases with unlabeled data. Finally, a new way of representing the static part of irregularly sampled light curves by a mixture of Gaussians is discussed. This represen- tation is more general than the extraction of features, as it allows the inclusion of photometric uncertainties and avoids the introduction of observational biases.