Incremental Clustering Using a Core-Based Approach

摘要

Clustering is a division of data into groups of similar objects, with respect to a set of relevant attributes (features) of the analyzed objects. Classical partitioning clustering methods, such as k-means algorithm, start with a known set of objects, and all features are considered simultaneously when calculating objects' similarity. But there are numerous applications where an object set already clustered with respect to an initial set of attributes is altered by the addition of new features. Consequently, a re-clustering is required. We propose in this paper an incremental, k-means based clustering method, Core Based Incremental Clustering (CBIC), that is capable to re-partition the objects set, when the attribute set increases. The method starts from the partitioning into clusters that was established by applying k-means or CBIC before the attribute set changed. The result is reached more efficiently than running k-means again from the scratch on the feature-extended object set. Experiments proving the method's efficiency are also reported.