2.2.2 Non-Linear Combinations
When we have a classification problem, and our learner outputs a discrete class label rather
than a real-valued number, a widely used combination rule is a majority vote among the
labels predicted by each member of the ensemble [7, 114]. Kittler and Alkoot [65] theoretically analysed the relationship between the sum and vote fusion strategies. They found
that when the errors in our estimate of the posterior probability of a class have a normal
distribution, sum always outperforms vote, whereas for heavy tail distributions, vote may
outperform sum; this was confirmed with empirical data. Kuncheva et al [72] derive theoretical upper bounds on the maximum achievable accuracy when using a majority voting
combination.
Bahler and Navarro [5] conduct a large empirical study into using different combination
rules. They found that when accuracy is approximately balanced across the estimators,
majority voting performs as well as any more complex combination rule. When accuracy
is imbalanced, majority vote tends to decrease in reliability, while more complex methods which take account of the individual performances, like Dempster-Schafer theory of
evidence [91] and Bayesian methods [2], retain their performance.
2.2.2 Non-Linear CombinationsWhen we have a classification problem, and our learner outputs a discrete class label ratherthan a real-valued number, a widely used combination rule is a majority vote among thelabels predicted by each member of the ensemble [7, 114]. Kittler and Alkoot [65] theoretically analysed the relationship between the sum and vote fusion strategies. They foundthat when the errors in our estimate of the posterior probability of a class have a normaldistribution, sum always outperforms vote, whereas for heavy tail distributions, vote mayoutperform sum; this was confirmed with empirical data. Kuncheva et al [72] derive theoretical upper bounds on the maximum achievable accuracy when using a majority votingcombination.Bahler and Navarro [5] conduct a large empirical study into using different combinationrules. They found that when accuracy is approximately balanced across the estimators,majority voting performs as well as any more complex combination rule. When accuracyis imbalanced, majority vote tends to decrease in reliability, while more complex methods which take account of the individual performances, like Dempster-Schafer theory ofevidence [91] and Bayesian methods [2], retain their performance.
การแปล กรุณารอสักครู่..
