Clarification regarding margin based uncertainty sampling.

[Kishaan Jeeveswaran]

Hi,

It is inferred from our research work on the efficiency of different active learning strategies for the task of automatic short answer grading that margin based sampling works better than other strategies when the feature space is high dimensional. However we couldn't find a theoretical reason to back this claim. Could someone help us infer the reason behind this? Could margin based uncertainty considering two classes' probabilities be connected to this in any way?

Thanks!