基于动态权重的Adaboost算法研究 *

针对Adaboost算法只能静态分配基分类器权重,不能自适应地对每个测试样本动态调整权重的问题,提出了一种基于动态权重的Adaboost算法。算法通过对训练样本集合进行聚类,并分析每个基分类器和每个类簇的适应性,进而为每个基分类器在不同类簇上设置不同权重,最终根据测试样本与类簇之间的相似性来计算基分类器在测试样本上的权重。在UCI数据集上的实验结果表明本文提出算法有效利用了测试样本之间的差异性,得到了比Adaboost算法更好的效果。

Research on Dynamic Weights Based Adaboost

Ensemble learning algorithms such as Adaboost combine base learners statically so that they cannot customize proper combinations of base learners for different testing samples. To deal with this drawback, this work proposes a new ensemble learning algorithm which weights base learner dynamically. In the training phase, training samples are grouped into clusters and each base leaner is associated with a weight that decided by the fitness between base learners and clusters. In the testing phase, the weights of base learners w.r.t a testing sample is calculated using the distance between it and the clusters. Experiment results on UCI datasets show that the proposed algorithm obtained better performance than traditional ensemble learning algorithms by exploring the diversity among testing samples.