Graph Meta Learning via Local Subgraphs

Prevailing methods for graphs require abundant label and edge information for learning. When data for a new task are scarce, meta learning can learn from prior experiences and form much-needed inductive biases for fast adaption to new tasks.