Stochastic Online Greedy Learning with Semi-bandit Feedbacks