Extending K-Means Clustering Algorithm

The K-Means algorithm for clustering has the drawback of always maintaining K clusters. This leads to ineffective handling of noisy data and outliers. Noisy data is defined as having little similarity with the closest cluster’s centroid. In K-Means a noisy data item is placed in the most similar cluster, despite this similarity is low relative to the similarity of other data items in the same cluster with the centroid. In part one this project, I have implemented an expanding version of the K-Means algorithm in an attempt to deal with noisy data more effectively. The idea is to first create a new cluster whenever a data item has a distance with the most similar cluster’s centroid beyond a threshold, and then place this data item in it. In other words, the number of clusters K is expandable to accommodate the clustering of data items which contains noisy data. In real applications of clustering, it is often the case that a data item belongs to more than a single class. For example, an article about the profitability of small market baseball teams has its place in both business and sports sections of a newspaper. In regular K-Means, each data item is placed in the most similar cluster, and this one only. In part two of this project, I have experimented with an overlapping version of K-Means. It places each data item in the clusters which have a similarity measure greater than or equal to a threshold. This allows a data item to be placed in as many clusters as necessary to deal with the clustering of data items which have overlapping classes.