Research Of Clustering Algorithms Based On Granular Computing

With the development and popularity of network technology, large amounts of data is full of people’s eyes, and people need to face massive data everyday. Basic storing and reading cannot satisfy the requirement already, people wish to find useful information from massive data. As a matter of fact, data mining is an effective technology which can find hidden information from massive data. As an important research area of data mining, clustering finds the regulations based on intra-class similarity and inter-class dissimilarity.As we know, clustering is an unsupervised learning method whose nature is equivalence partition for the universe. The partition results are that intra-class similarity is greater than the given threshold and inter-class similarity less. In fact, traditional clustering algorithm is hard to deal with the fuzzy and uncertain data, insensitive to abnormal information, always unstable, too dependent on input parameters and slow when processing large data sets. For all its faults, this paper proposes a fusion model clustering algorithm which combines three models of granular computing: fuzzy set, rough set and quotient space.First, this paper introduces the basic concepts of granular computing and analyzes clustering algorithms for single model, fuzzy quotient space and fuzzy rough set. Based on the original model clustering algorithm, the fusion model clustering algorithm is developed by combining advantages of each model. Specifically, fuzzy clustering algorithm can deal with fuzzy information to make the results more consistent with the actual; rough set clustering algorithm can get optimal parameters resulting in more accurate results and more stable algorithm; quotient space clustering algorithm can get different clustering results with different granularity and find the optimal one, which is in accordance with the process of human problem solving.In order to deal with mixed valued data, this paper use unified dissimilarity formula to compute the distance of fuzzy quotient space. Then calculate the threshold of each cluster by shadowed sets theory, obtain positive region and boundary region of each cluster according to the threshold and apply approximation weight method to find centers of clusters. The centers will be taken as the input of the next iteration, then keep iterating until the clustering results are stable. In theory, the proposed method needs few parameters and computes more accurate distance between mixed valued samples. Since each iteration is on the basis of the last clustering results, the algorithm has shorter execution time and is sensitive to abnormal information. Experiments show that the fusion model clustering algorithm owns more excellent performance, which is also consistent with theoretical analysis.