Abstract:
|
In this paper, we design the transfer learning methods for the high-dimensional quantile regression (QR), utilizing the different QR models built on the source populations to improve the performance of QR on the target population. Given certain informative sources, the oracle transfer learning quantile regression is carried out with its l2-estimation error bound and l2-prediction error bound. When the informative sources are unknown, to make inferences on the index of informative sources, a cross-validation-based informative source detection procedure is introduced. The group theoretically proves that under mild conditions, the detection is consistent, and further, the transfer learning QR enjoys a sharper convergence rate than the QR with only the target dataset while circumvents the negative learning problem. Groups of simulations verify this improvement of performance. Real-data analysis on the hard landing risk detection for flight safety is implemented with the proposed transfer learning method, proved with better performance by the cross-validation.
|