特征选择的和声模糊聚类研究与应用

摘要

Three methods are adopted to achieve a better clustering quality. Considering the different influences of each dimen-sion attribute of data on the clustering effect, statistic method is used to weight each dimension to select feature. Some improve-ments are carried out for the probability of harmony search algorithm and combine fuzzy clustering algorithm with harmony search to rapidly find the optimal cluster centers. The iterative method is used to test clustering quality to get the best number of cluster center. The proposed algorithm is applied to parallel computing performance analysis to distinguish and identify the per-formance bottleneck category of various processors during parallel program running. Experimental results show that the pro-posed clustering algorithm outperforms other similar algorithms. This performance analysis method can improve the operating efficiency of the parallel program.%采取了3种必要的措施提高了聚类质量：考虑到各维数据特征属性对聚类效果影响不同，采用了基于统计方法的维度加权的方法进行特征选择；对于和声搜索算法的调音概率进行了改进，将改进的和声搜索算法和模糊聚类相结合用于快速寻找最优的聚类中心；循环测试各种中心数情况下的聚类质量以获得最佳的类中心数。该算法被应用于并行计算性能分析中，用于识别并行程序运行时各处理器运行性能瓶颈的类别。实验结果表明该算法较其他算法更优，这样的性能分析方法可以提高并行程序的运行效率。