Label space driven feature space remapping

Abstract

The elicitation of labeled training data from physical sources is a primary bottleneck that limits the applicability of traditional supervised learning algorithms. Transfer learning algorithms overcome the limitation for a target domain where training data is scarce or absent by leveraging labeled data from related auxiliary domains. The proposed novel transfer learning frameworks yield a sparse and linear transformation to bridge domains with heterogeneous features without relying on instance or feature correspondences. Extensive experiments are conducted on real-world heterogeneous transfer tasks with diverse datasets of varying dimensions and sparsity to investigate the effectiveness of the proposed approaches over other baseline and state of the art transfer approaches.