Semi-supervised Projected Clustering

Speaker: Kevin YIP

Abstract

Projected clustering has found a number of potential applications in areas such as bioinformatics, time-series analysis and document categorization. However, most existing projected clustering algorithms are unable to detect clusters of very low dimensionalities (e.g. 5% of dataset dimensionality) without placing some constraints on the cluster properties or requiring users to input some parameter values that are hard for them to obtain. Inspired by some recent works in the machine learning community, we propose the use of a small amount of domain knowledge (which is usually available in some applications) to guide the clustering process. According to some preliminary experimental results, our new algorithm has a better performance than some projected and non-projected algorithms even no external inputs are supplied, and a steady performance improvement is observed when the amount of external inputs is increased. In the talk I will describe the semi-supervised projected clustering problem, our new algorithm, and also some encountered difficulties and possible extensions of the study.

Referred Papers