Clustering large data with uncertainty

A new algorithm is designed for handling fuzziness while mining large data. A new novel cost function weighted by fuzzy membership, is proposed in the framework of CLARANS. A new scalable approximation to the maximum number of neighbors, explored at each node, is developed; thus reducing the computational time for large data while eliminating the need for user-defined (heuristic) parameters in the existing equation. The goodness of the generated clusters is evaluated in terms of Xie–Beni validity index. Results demonstrate the superiority of the proposed algorithm, over both synthetic and real data sets, in terms of goodness of clustering. It is interesting to note that our algorithm always converges to the globally best values at the optimal number of partitions. Moreover compared to existing fuzzy algorithms, FCLARANS without scanning the whole dataset, searching small number of neighbors, is able to handle the uncertainty due to overlapping nature of the various partitions. This is the main motivation of fuzzification of the algorithm CLARANS.

Variation of threshold Ncross with the number of clusters c. Figure optionsDownload as PowerPoint slideHighlights
► A fuzzified version of the algorithm CLARANS is designed.
► The cost function is weighted by fuzzy membership in the framework of CLARANS.
► A new scalable approximation to the maximum number of neighbors, explored at each node.