Self-adaptive weighted majority vote algorithm based on entropy
Rui Li, Xiaodan Wang · 2017
Based on the novel idea of assigning different weight for same sample by different classifier according to the classification ability and assigning different weight for different sample by same classifier according to the separable degree of sample respectively, a self-adaptive weighted vote algorithm based on entropy is proposed. Probability Multi-class Relevance Vector Machine (PRVM_OVO) is extended with approach `pairwise coupling' based on the basic RVM model, and three different PRVM_OVO are used to classify different feature samples, then entropy calculated by the posterior probability of different PRVM_OVO is used to assign weight adaptively, low weight is assigned to sample with high entropy, at last weighted majority vote strategy is used to fusion different classifier's results to get the final recognition results. Experiment results based on UCI datas show the efficiency of the proposed method.