谱聚类中特征向量的Bagging选取方法

谱聚类算法中用亲和矩阵特征值最大的k个特征向量并不总是能有效地发现数据集的结构。为了选取较好特征向量,提出了一种特征向量的Bagging选取算法。以成对约束计分方法为评价标准,对特征向量进行评价并选出较好的特征向量,将多次选择的特征向量进行Bagging集成(Bootstrap aggregating),得出k个特征向量的组合。该算法能够较好地选取出特征向量,根据UCI实验数据集的测试,证实该算法对测试数据集可以得出较好的预测结果。

Eigenvector selection in spectral clustering based on Bagging

For the spectral clustering algorithm, the largest k eigenvectors of the affinity matrix derived from the dataset were not always able to find the structure of dataset effectively. An eigenvector selection algorithm in spectral clustering based on Bagging was proposed in order to select better eigenvectors. The eigenvectors were evaluated by pairwise constraints score. First, some eigenvectors were ranked according to their constraint scores, and then the suitable eigenvectors were selected from the ranking list, finally the optimal combination of k eigenvectors was obtained by Bagging based ensemble algorithm. The better eigenvectors could be achieved. Experimental results on UCI benchmark datasets showed that this algorithm could gain satisfactory prediction results.