Towards making co-training suffer less from insufficient views