Mixed-Attribute Clustering and Weighted Clustering

Speaker: Ken YIU

Abstract Clustering may not be meaningful in high dimensional spaces because of the presence of irrelevant attributes. Projected clustering algorithms have been developed to find the relevant attributes for the clusters during cluster formation. However, it is possible that not all relevant attributes have the same relevance to the clusters. Therefore, we attempt to extend projected clustering algorithms to weighted clustering algorithms. Unlike projected clustering, weighted clustering considers all the dimensions. Weights are used to represent the relevances of each dimension to each cluster. In the presentation, we will study how to define the weights. We will adapt PROCLUS (a projected clustering algorithm) to a weighted clustering algorithm. Also, three methods for computing the weights will be presented.

In addition, there have been few clustering algorithms on mixed attributes but most of the real datasets have both numeric and categorical attributes. Combined information from mixed attributes may be useful for clustering. Two papers for clustering categorical data will be briefly introduced. Then, we will study how to extend them to mixed attributes.

Referred Papers