B0883
Title: Robust transfer learning in high-dimensional GLM via gamma-divergence
Authors: Fuzhi Xu - Xiamen University (China)
Yaqing Xu - Shanghai Jiao Tong University School of Medicine (China) [presenting]
Shuangge Ma - Yale University (United States)
Qingzhao Zhang - Xiamen University (China)
Abstract: Outlying observations and even data contamination often occur in practice due to high-dimensional sparsity. Robustness against outliers and contamination based on the divergence has been widely adopted. With the rapid growth in the volumes of high-dimensional data, learning from multiple sources of evidence is desired. Transfer learning can improve the performance of target models by transferring information from source datasets. Yet, multiple sources of information, introducing outlying observations and even contamination, may lead to biased estimation and misleading inference. A robust transfer learning approach is proposed based on the minimum gamma-divergence under a generalized linear model (GLM) framework for high-dimensional data. Using a robust algorithm-free transferable source detection scheme, the proposed approach identifies informative sources and avoids negative transfer of learning. The consistency properties and estimation error bounds under high dimensionality are rigorously established. A computational algorithm is developed based on proximal gradient descent for transferring and debiasing steps. Simulation demonstrates the superior and competitive performance of the proposed approach in selection and prediction/classification. Analysis of genetic data on breast cancer and glioblastoma confirms its practical usefulness.
