Incentivizing High Quality User Contributions: New Arm Generation in Bandit Learning