DiReT: An Effective Discriminative Dimensionality Reduction Approach for Multi-Source Transfer Learning

Learning invariant features while the distribution of training (source domain) and test (target domain) sets are different is crucial. However, most of the dimensionality reduction methods perform poorly in facing with domain shift problem either in original or latent spaces. In this paper we introduce DiReT, a Discriminative Dimensionality Reduction approach for multi-source Transfer learning, that aims at constructing a latent space across domains in a semi-supervised manner. Our main contribution is to reduce the drift in distributions across domains and concurrently preserve the separability between classes. DiReT by establishing a bridge between source and target domains guarantees the knowledge transformation along different domains. Empirical evidences indicate that DiReT manages to improve substantially over dimensionality reduction methods, especially with extracting more features from multiple domains. We evaluate DiReT against other well-known dimensionality reduction and transfer learning methods on three synthetic and two real world datasets.