Date on Master's Thesis/Doctoral Dissertation

5-2011

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Committee Chair

Frigui, Hichem

Author's Keywords

Clustering; Kernel learning; Fuzzy; Pattern recognition; Information retrieval

Subject

Pattern recognition systems; Cluster analysis; Data mining

Abstract

Despite the large number of existing clustering methods, clustering remains a challenging task especially when the structure of the data does not correspond to easily separable categories, and when clusters vary in size, density and shape. Existing kernel based approaches allow to adapt a specific similarity measure in order to make the problem easier. Although good results were obtained using the Gaussian kernel function, its performance depends on the selection of the scaling parameter. Moreover, since one global parameter is used for the entire data set, it may not be possible to find one optimal scaling parameter when there are large variations between the distributions of the different clusters in the feature space. One way to learn optimal scaling parameters is through an exhaustive search of one optimal scaling parameter for each cluster. However, this approach is not practical since it is computationally expensive especially when the data includes a large number of clusters and when the dynamic range of possible values of the scaling parameters is large. Moreover, it is not trivial to evaluate the resulting partition in order to select the optimal parameters. To overcome this limitation, we introduce two new fuzzy relational clustering techniques that learn cluster dependent Gaussian kernels. The first algorithm called clustering and Local Scale Learning algorithm (LSL) minimizes one objective function for both the optimal partition and for cluster dependent scaling parameters that reflect the intra-cluster characteristics of the data. The second algorithm, called Fuzzy clustering with Learnable Cluster dependent Kernels (FLeCK) learns the scaling parameters by optimizing both the intra-cluster and the inter-cluster dissimilarities. Consequently, the learned scale parameters reflect the relative density, size, and position of each cluster with respect to the other clusters. We also introduce semi-supervised versions of LSL and FLeCK. These algorithms generate a fuzzy partition of the data and learn the optimal kernel resolution of each cluster simultaneously. We show that the incorporation of a small set of constraints can guide the clustering process to better learn the scaling parameters and the fuzzy memberships in order to obtain a better partition of the data. In particular, we show that the partial supervision is even more useful on real high dimensional data sets where the algorithms are more susceptible to local minima. All of the proposed algorithms are optimized iteratively by dynamically updating the partition and the scaling parameter in each iteration. This makes these algorithms simple and fast. Moreover, our algorithms are formulated to work on relational data. This makes them applicable to data where objects cannot be represented by vectors or when clusters of similar objects cannot be represented efficiently by a single prototype. Our extensive experiments show that FLeCK and SS-FLeCK outperform existing algorithms. In particular, we show that when data include clusters with various inter-cluster and intra-cluster distances, learning cluster dependent kernel is crucial in obtaining a good partition.

Share

COinS