基于粒计算的多尺度聚类尺度上推算法

Upscale algorithm of multi-scale clustering based on granular computing

多尺度科学在数据挖掘领域的研究多见于图像和空间数据挖掘,对一般数据的多尺度特性研究较少。传统聚类算法只在单一尺度上进行,无法充分挖掘蕴藏在数据中的知识。引入粒计算思想,进行普适的多尺度聚类方法研究,对数据进行多层次、多角度分析,实现一次挖掘,多次应用。首先,介绍粒计算相关知识;然后,提出多尺度聚类尺度上推算法UAMC(upscaling algorithm of multi-scale clustering),以簇为粒子,簇心为粒子特征进行尺度转换,利用斑块模型得到大尺度知识,避免二次挖掘带来的资源浪费。最后,利用UCI公用数据集和H省全员人口真实数据集对算法性能进行实验验证,结果表明算法在准确性上优于K-means等基准算法,是有效可行的。

Research of multi-scale scientific mainly focuses on space or image data in the field of data mining, while paying less attention to multi-scale features of general data. Traditional clustering algorithms are implemented based on single scale, which are not able to discover potential knowledge in data. This paper carried out a study of methods on universal multi-scale clustering with the introduction of granular computing, for the purpose of muhilayer and multi-angle of data analysis and single- mining-multiple-using. First of all, this paper described knowledge related to granular computing. Then, it proposed an algorithm called UAMC ,with clusters as granularity and clustering centers as feature of granularity to scale conversion, obtaining know- ledge of large scale based on mosaic upscaling scheme, for fear of resource waste due to secondly mining. At last, experimental results on datasets from UCI and H province indicate that UAMC algorithm outperforms benchmark algorithms such as K-means in accuracy. Meanwhile, UAMC algorithm is verified to be effective and feasible through the experiments.