Difficulty-aware deisolated affinity propagation based collaborative learning via dynamic constrained multi-objective optimization