Abstract: Ensemble learning is widely accepted and used in machine learning. This paper
proposes a multi-class ensemble learning algorithm named AdaBoost belief. The algorithm
improves AdaBoost·SAMME by attaching weights to classes in every weak classifier. These
weights, called class beliefs, are computed based on class accuracy collected in each round
of the iteration. We compare the algorithm with AdaBoost·SAMME in many aspects including
learning accuracy, generalization ability, and theory support. Experimental results
indicate that the proposed method has a competitive learning ability and high prediction
accuracy in Gaussian sets, several UCI sets, anda number of log-based intrusion detection
applications. When the class number increases so that prediction of classes becomes
more difficult, the prediction error rate of the proposed algorithm increases slower than
AdaBoost·SAMME.

Key words: ensemble learning, multi-class, class belief, class weight, AdaBoost·SAMME