Transfer Sparse Machine: Matching Joint Distribution by Subspace Learning and Classifier Transduction

摘要

Transfer learning problem aims at matching the joint distributions of the source and the target datasets so that the model learned from the source dataset can be applied to the target dataset. Unfortunately, the joint distribution of the database may be very hard to estimate in many applications. Since the joint distribution can be written as the product of the marginal and the conditional distributions, we propose the TSM, which tries to match the latter two distributions respectively, instead of directly matching the joint distributions. The proposed TSM consists of two parts: a feature learning part which matches the marginal distributions by learning a shared feature space, and a classifier training part which matches the conditional distributions by training an adaptive classifier in the shared feature space. Comprehensive experiments prove that the superior performance of the TSM on several transfer learning datasets. And the improvements are 12.86% on the USPS/MNIST dataset and 9.01% on the PIE1/PIE2 dataset compared to the best baseline.