A0600
Title: Bayesian transfer learning with multiple auxiliary datasets
Authors: Donatello Telesca - UCLA (United States) [presenting]
Abstract: Transfer learning is considered in the context of high dimensional linear and generalized linear models. While Bayesian inferential methods are naturally suited for TL, some care is needed in the construction of sparsity-inducing priors to mitigate the effects of possible negative transfers. When multiple auxiliary datasets are available to inform a target task, we show how the combination of regularization priors with standard Bayesian model selection can prove effective in the identification of an informative auxiliary set while accounting for uncertainty in the selection of important covariates. A direct parametrization through sparse contrasts allows for fine-tuning of the level of borrowing and desired levels of model compatibility. Finally, parallels are discussed with related techniques in Bayesian meta-analysis and several classes of power/power-like priors.
