On the Minimax Regret for Online Learning with Feedback Graphs