Proactive Transfer Learning for Heterogeneous Feature and Label Spaces.

We propose a framework for learning new target tasks by leveraging existing heterogeneous knowledge sources. Unlike the traditional transfer learning, we do not require explicit relations between source and target tasks, and instead let the learner actively mine transferable knowledge from a source dataset. To this end, we develop (1) a transfer learning method for source datasets with heterogeneous feature and label spaces, and (2) a proactive learning framework which progressively builds bridges between target and source domains in order to improve transfer accuracy. Experiments on a challenging transfer learning scenario (learning from hetero-lingual datasets with non-overlapping label spaces) show the efficacy of the proposed approach.