Abstract

Researcher(s): Sumin Lee

Link prediction is a study that automatically finds and fills in missing facts among existing links. Since it is impractical to obtain appropriate training data for newly added links, zero-shot learning that recognizes unseen classes without training data is preferred.

Therefore, implemented a zero-shot link prediction model that infers embedding for unseen links using a generative model learned on seen links, and apply meta pseudo label to improve the quality of inferred embedding by distillation and feedback between the teacher model and the student model, to predict better quality pseudo data.