D. Luo, X. Wu, and H. Chi (PRC)
Outlier, Ensemble learning, AdaBoost, Unit-weighting
Statistical ensemb le learning methods have turned to be effective way to improve accuracy of a learning system [1][2][3]. However, traditional ensemble methods will perform worse if there exist a lot of outliers [4][5][6]. The data selection strategies in traditional ensemb le methods are based on the sample-weighting mechanism and lead to the outlier problem serious. In this paper, a new data selection strategy that based on the unit-weighting mechanism is proposed, where the weight of a sample is no longer determined only by the sample itself, it will also be influenced by the unit the sample belongs to. The simulation results on speaker identification using KING benchmark database show that the proposed data selection strategy is effective in dealing with the outliers and successful in improving the identification accuracy.
Important Links:
Go Back
