Distant Transfer Learning via Deep Random Walk

Transfer learning, which is to improve the learning performance in target domain by leveraging useful knowledge from source domain, often requires that those two domains are very close, limits its application scope. Recently, distant transfer has been studied between or even totally unrelated via unlabeled auxiliary act as a bridge spirit of human transitive inference completely concepts can be connected through gradual transfer. In this paper, we study proposing DeEp Random Walk basEd distaNt (DERWENT) method. Different existing models implicitly identify path and instances instances, proposed DERWENT model explicitly learn such paths deep random walk technique. Specifically, based on sequences identified technique data graph where have no direct connection, enforces adjacent points sequence similar, makes ending point represented other same sequence, considers weighted classification losses data. Empirical studies several benchmark datasets demonstrate algorithm yields state-of-the-art performance.