Multitask Bandit Learning through Heterogeneous Feedback Aggregation