An investigation of K-means clustering to high and multi-dimensional biological data

The K-means clustering algorithm has been intensely researched owing to its simplicity of implementation and usefulness in the clustering task. However, there have also been criticisms on its performance, in particular, for demanding the value of K before the actual clustering task. It is evident from previous researches that providing the number of clusters a priori does not in any way assist in the production of good quality clusters. Our investigations in this paper also confirm this finding. The objective of this paper is to investigate further, the usefulness of the K-means clustering in the clustering of high and multi-dimensional data by applying it to biological sequence data. The squared Euclidean distance and the cosine measure are used as the similarity measures. We use the silhouette validity index first to show that K-means algorithm is not suitable for clustering high and multi-dimensional biological data irrespective of the distance or similarity measure employed. A preprocessor scheme is then added to the Kmeans algorithm. The scheme is used to automatically initialize a suitable value of K prior to the execution of the K-mean algorithm. Central to the preprocessor is the average silhouette value of the clusters. Our investigation suggests that the use of the silhouette value in the preprocessor improves the quality of clusters significantly for the biological datasets considered. Furthermore, we suggest a scheme which maps the high dimensional data into low dimensions. We have then shown that the K-means algorithm with preprocessor produces good quality, compact and well-separated clusters of the biological data mapped in low dimensions. For the purpose of clustering we conduct a character-to-numeric conversions to transform the nucleic/amino acids symbols to numeric values.