The Application Of Feature Weight Learning In Unsupervised Clustering

Cluster analysis is one of the means of multivate analysis and is widely used in pattern recognition, data mining and decision analysis etc. Now the popular clustering algorithm is transitive closure clustering algorithm and Fuzzy c means clustering algorithm. Basically there is fuzziness in the procedure of clustering. In this paper, each feature is considered to have an importance degree which is called feature weight. When one feature weight equals to zero, then the feature should be deleted in clustering. With feature weight increasing, the feature is more important in clustering. This paper proposes an approach to feature weight learning which is based on the gradient-decent technique. It shows that an appropriate assignment to feature- weights can reduce the fuzziness and improve the performance of clustering algorithm. Experiments on some UCI databases and web usage mining demonstrate the improvements of performance of clustering algorithm.