We analyze imbalanced training loss, showing that gradient descent dynamics can gradually reduce bias and recover minority-specific features with longer training.

We present new tests for detecting a community in a random bipartite graph and prove that they are minimax optimal.