Application Of Multi-task Learning Method Based On Lasso In The Stellar Parametrization

Multi-task learning approaches have attracted the increasing attention in machine learning, computer vision, and artificial intelligence. Learning multiple related tasks simultaneously is better than learning each task independently, by utilizing tasks’ relevance.Lasso method can reduce the data’s dimension, extract the feature vector, reduce the amount of the data, eliminate the noise’s interference, and improve the parameters’ accuracy. An efficient multi-task Lasso regression algorithm is adopted in this paper to estimate the stellar spectra physical parameters: the surface effective temperature lg Teff、the acceleration of gravity lg g、the chemical abundance [Fe/H]. The experiments’ results are better than the related researches’ methods and the single-task Lasso regression method. The three parameters’ accuracy are all improved, especially for the lg g and [Fe/H]. In experiments, we changed the resolution of the spectrum and applied different levels of SNR noise to the spectrum, so as to analysis the estimation of the stellar spectra’s physical parameters comprehensively. Results show that the model is effected both on the resolution and the noise. But the influence of the noise is larger than resolution’s. In general, the multi-task Lasso regression method is suitable for the automatic estimation of the stellar spectra’s physical parameters. It calculates fast, consumes less time,operates simply. What’s more, it can improve the overall accuracy of the model.