Affiliations: Department of Computer Science, Renmin University of China, Beijing, China | Department of Computer & Information Science, Indiana University Purdue University Indianapolis, Indianapolis, IN, US | Department of Mathematic Science, Indian University Purdue University Indianapolis, Indianapolis, IN, US
Note: [] This work is partially funded by the National Natural Science Foundation of China under Grant No.61170012.
Note: [] Corresponding author. Yuni Xia, Department of Computer & Information Science, Indiana University Purdue University Indianapolis, 723 W Michigan St, Indianapolis, IN, US. E-mails: [email protected]; [email protected] (Biao Qin); [email protected] (Fang Li); [email protected] (Jiaqi Ge).
Abstract: Real world applications as sensor networks and RFID networks usually generate data with uncertainty. Data uncertainty comes from many sources, as measurement errors, limited precision, data aggregation and so on. Classical data mining applications need to be modified and extended for uncertain data; otherwise, their performances might be dramatically downgraded by data uncertainty. In this paper, we define an uncertain data model for both numerical and categorical uncertain data, and propose a new Expectation-Maximization based algorithm (EMU) for clustering uncertain data. This approach is well designed to find the distribution parameters that maximize model qualities based on uncertain data, therefore correctly identify the clusters. Our clustering algorithm can process both numeric and categorical uncertain data. In our experiments, we use both synthetic and real data sets to evaluate the effectiveness and robustness of the proposed algorithm.
