A0712
Title: Deep representation transfer learning for partially linear models
Authors: Baihua He - University of Science and Technology (China) [presenting]
Abstract: Transfer learning has achieved many successes in practical applications. Some issues remain unexplored, especially the combination of parameter interpretability and model flexibility. A novel deep neural network-based transfer learning approach is presented within a partially linear model to enhance model performance using heterogeneous source domain data. The proposal addresses the challenges of high-dimensional and non-linear data, offering both high prediction accuracy and improved interpretability of primary parameters. Distinct from existing statistical methods, the framework facilitates positive transfer under certain domain heterogeneity, maintaining robustness against over-fitting through data augmentation. The consistency of representation learning, asymptotic normality of the primary parameters, and semi-parametric efficiency are established. The promising performance of the proposal is demonstrated through simulations and empirical validations on the house renting price study
