Image and Video Analysis

Using k-nearest neighbor and feature selection as an improvement to hierarchical clustering

Methods and Applications of Artificial Intelligence, Springer, Volume 3025, pp.0-0, 2004.

Clustering of data is a difficult problem that is related to various fields and applications. Challenge is greater, as input space dimensions become larger and feature scales are different from each other. Hierarchical clustering methods are more flexible than their partitioning counterparts, as they do not need the number of clusters as input. Still, plain hierarchical clustering does not provide a satisfactory framework for extracting meaningful results in such cases. Major drawbacks have to be tackled, such as curse of dimensionality and initial error propagation, as well as complexity and data set size issues. In this paper we propose an unsupervised extension to hierarchical clustering in the means of feature selection, in order to overcome the first drawback, thus increasing the robustness of the whole algorithm. The results of the application of this clustering to a portion of dataset in question are then refined and extended to the whole dataset through a classification step, using k-nearest neighbor classification technique, in order to tackle the latter two problems. The performance of the proposed methodology is demonstrated through the application to a variety of well known publicly available data sets.

[ Bibtex ] [ PDF ]