Weighted likelihood transfer learning for high-dimensional generalized linear models
Zhaolei Liu, Lu Lin · Statistics · 2024
To simultaneously improve parameter estimation and variable selection for a target model by the auxiliary information from source models, a weighted likelihood transfer learning (WL-TL), together with a l1-penalty, is proposed for high-dimensional generalized linear models. To implement the transfer learning, the relevant techniques, including iterative algorithm and the choice of weight, are suggested. The methodology is computational simple, without need for the bias-correction used in the existing literature of parameter-transfer learning. The theoretical properties such as the quadratic error bound of the parameter estimator and the estimation consistency are established. A specific weight selection method based on the Bayesian decision theory has been proposed and studied. Comprehensive simulation experiments and real data analyzes are conducted to further illustrate the performance of the new method.