Informative regularization for a multi-layer perceptron RR Lyrae classifier under data shift

In recent decades, machine learning has provided valuable models and algorithms for processing extracting knowledge from time-series surveys. Different classifiers have been proposed performed to an excellent standard. Nevertheless, few papers tackled the data shift problem in labeled training sets, which occurs when there is a mismatch between distribution set testing set. This drawback can damage prediction performance unseen data. Consequently, we propose scalable easily adaptable approach based on informative regularization ad-hoc procedure mitigate during of multi-layer perceptron RR Lyrae classification. We collect ranges characteristic features construct symbolic representation prior knowledge, was used model regularizer component. Simultaneously, design two-step back-propagation algorithm integrate this into neural network, whereby one step applied each epoch minimize classification error, while another ensure regularization. Our defines subset parameters (a mask) loss function. handles forgetting effect, stems trade-off these functions (learning versus expert knowledge) training. Experiments were conducted using recently shifted benchmark sets stars, outperforming baseline by up 3% through more reliable classifier. method provides new path incorporate artificial networks manage underlying problem.