Improving the performance of photometric regression models via massive parallel feature selection

Regression tasks are common in astronomy, for instance, the estimation of the redshift or the metallicity of galaxies. Generating regression models, however, is often hindered by the heterogeneity of the available input catalogs, which leads to missing data and/or features of differing explanatory power. In this work, we show how simple but effective feature selection schemes from data mining can be used to significantly improve the performance of regression models for photometric redshift and metallicity estimation (even without any particular knowledge of the input parameters' physical properties). Our framework tests huge amounts of possible feature combinations. Since corresponding (single-core) implementations are computationally very demanding, we make use of the massive computational resources provided by nowadays graphics processing units to significantly reduce the overall runtime. This renders an exhaustive search possible, as we demonstrate in our experimental evaluation. We conclude the work by discussing further applications of our approach in the context of large-scale astronomical learning settings.