These are important questions to ask, for two reasons. The first is from the viewpoint
of NC as a stand-alone tool: anything that helps us understand its behaviour or configure
its parameters, is useful. The second reason is from the viewpoint of gaining a better
understanding of ensemble classification error “diversity”. NC is a technique that was
claimed to explicitly enforce this type of diversity; in his thesis Liu [85] shows that a lower
correlation between classification errors is observed when using NC. Understanding why
NC works should clarify, at least partially, what “diversity” means. If we can formalise
“diversity”, and then control it, we may be able to engineer better performing ensemble
systems. A third thesis question is therefore “Can we identify a good way to understand
classifier error diversity, and can we explicitly control it?”.
These are important questions to ask, for two reasons. The first is from the viewpointof NC as a stand-alone tool: anything that helps us understand its behaviour or configureits parameters, is useful. The second reason is from the viewpoint of gaining a betterunderstanding of ensemble classification error “diversity”. NC is a technique that wasclaimed to explicitly enforce this type of diversity; in his thesis Liu [85] shows that a lowercorrelation between classification errors is observed when using NC. Understanding whyNC works should clarify, at least partially, what “diversity” means. If we can formalise“diversity”, and then control it, we may be able to engineer better performing ensemblesystems. A third thesis question is therefore “Can we identify a good way to understandclassifier error diversity, and can we explicitly control it?”.
การแปล กรุณารอสักครู่..
