Targeted Learning of Heterogeneous Sources by Informative Feature Sharing

Music City Center 

Transfer learning has been proven useful for leveraging information from multiple similar source datasets to enhance the performance of the target model. A fundamental challenge in transfer learning is avoiding negative transfer when there is heterogeneity among the sources and between the source and target datasets. Traditional methods are typically based on identifying informative sources. This creates a binary all-in or all-out decision, potentially resulting in the loss of useful information. In this paper, we introduce Targeted-IFS, a new transfer learning framework for high-dimensional Generalized Linear Models (GLMs) under heterogeneous sources. To avoid negative transfer and ensure effective transfer of useful information from sources, Targeted-IFS employs a pre-transfer debiasing step to correct estimates of selected informative features across all sources, rather than selecting the informative sources. We theoretically show that the Targeted-IFS method avoids negative transfer, achieving a convergence rate no worse than the classical LASSO using only target data, regardless of source heterogeneity. Simulations confirm its robustness to complex source heterogeneity and imp

Generalized linear model

heterogeneity

informative support

negative transfer

robust transfer learning 

Biometrics Section