Performance Evaluation Results of evolutionary clustering algorithm star for clustering heterogeneous datasets

The data was collected from the written Java codes by the authors, and Weka packages for executing ECA* on 32 heterogenous and multi-featured datasets against its counterpart algorithms (KM, KM++, EM, LVQ, and GENCLUST++). Each of these algorithms was run thirty times on each of the 32 benchmarking dataset problems to evaluate the performance of each of them in three tests. First, evaluating the performance of ECA* against its counterpart algorithms using internal and external measuring criteria. Second, analysing the statistical performance of ECA* compare to its counterpart algorithms based on their best solution (best), the worst solution (worst), and the mean solution (average) for intraCluster distance, interCluster distance, and execution time. Final, proposing a performance framework to investigate how sensitive the performance of these algorithms on different dataset features, such as cluster overlap, number of clusters, cluster dimensionality, and cluster structure, and cluster shape.

本数据集源自作者编写的Java代码，以及用于在32个异构多特征数据集上执行ECA*算法，并与对比算法（KM、KM++、EM、LVQ及GENCLUST++）开展对比的Weka工具包。本次研究针对这32个基准数据集问题，将所有对比算法各运行30次，以通过三项测试评估各算法的性能表现：第一项测试采用内部与外部评价准则，对比ECA*与其余对比算法的性能；第二项测试基于各算法在簇内距离、簇间距离与运行时间上的最优解、最差解及平均解，统计分析ECA*与对比算法的性能表现；最后，本研究提出一种性能分析框架，用于探究这些算法的性能对不同数据集特征的敏感程度，涵盖簇重叠度、簇数量、簇维度、簇结构及簇形状。