基于重采样策略的选择性谱聚类集成学习算法

提出了一种新的基于双重采样的选择性集成学习算法。针对集成学习要求学习器个体的差异性分布在样本空间的不同部分,对得到的聚类个体学习器输出进行重采样,以此来计算聚类个体的差异性。针对集成学习要求得到的个体学习器具有一定的精确性,对所有得到的学习器个体集合进行重采样来评估聚类个体精确性。在此基础上选择出集成学习所需的个体集合。以谱聚类算法作为基学习器,用聚类集成策略部分解决了谱聚类算法存在的尺度参数敏感问题,在UCI数据集上的仿真实验验证了算法的有效性。

Resampling-based Spectral clustering ensemble selection

An novel clustering ensemble selection approach is proposed in this paper. The diversity, which must be spreaded in the different part of the samples space, is required in ensemble learning.The output of the component clusteirngs is resampled to get the diversity of components. In ensemble learning, accuracy is also needed. The set of component clustering is resampled to assess the accuracy of components. Based on the resampling technique, we can select the appropriate individuals to construct ensemble learning system. The new proposed Spectral Clustering(SC) is exploited as the base learner and the sensitivity of scaling parameters of SC is partially soloved via ensemble strategy. The experimental results on UCI data demonstrate that the proposed algorithm is effective.