(Bensusan & Kalousis, 2001) have recently investigated Meta-Learning in a
similar setting, using features extended from STATLOG, histograms of numeric
attributes and landmarking using seven learners, four of which are also used by
us as base classifiers. They considered three meta-learners. They found that
landmarking is superior to predict errors of classifiers which was confirmed in our
experiments. However, they found that landmarking is not useful to generate
good rankings for classifiers – no meta-learner is able to perform better than a
default strategy. Our results indicate that for predicting significant differences,
landmarking features are also less useful. Mean absolute errors are given, but
no statistical correlation measure can be found for the regression experiments.
Some regression rules from Cubist are shown and interpreted, even though they
state that ranking determine from Cubist’s error prediction always perform
worse than the default.
(Bensusan & Kalousis, 2001) have recently investigated Meta-Learning in asimilar setting, using features extended from STATLOG, histograms of numericattributes and landmarking using seven learners, four of which are also used byus as base classifiers. They considered three meta-learners. They found thatlandmarking is superior to predict errors of classifiers which was confirmed in ourexperiments. However, they found that landmarking is not useful to generategood rankings for classifiers – no meta-learner is able to perform better than adefault strategy. Our results indicate that for predicting significant differences,landmarking features are also less useful. Mean absolute errors are given, butno statistical correlation measure can be found for the regression experiments.Some regression rules from Cubist are shown and interpreted, even though theystate that ranking determine from Cubist’s error prediction always performworse than the default.
การแปล กรุณารอสักครู่..
