Bandit Learning for Diversified Interactive Recommendation