Fastmap projection for high-dimensional data: A cluster ensemble approach

High-dimensional data with many features present a significant challenge to current clustering algorithms. Sparsity, noise, and correlation of features are common properties of high-dimensional data. Another essential aspect is that clusters in such data often exist in various subspaces. Ensemble clustering is emerging as a leading technique for improving robustness, stability, and accuracy of high-dimensional data clusterings. In this paper, we propose FastMap projection for generating subspace component data sets from high-dimensional data. By using component data sets, we create component clusterings and provides a new objective function that ensembles them by maximizing the average similarity between component clusterings and final clustering. Compared with the random sampling and random projection methods, the component clusterings by FastMap projection showed high average clustering accuracy without sacrificing clustering diversity in synthetic data analysis. We conducted a series of experiments on real-world data sets from microarray, text, and image domains employing three subspace component data generation methods, three consensus functions, and a proposed objective function for ensemble clustering. The experiment results consistently demonstrated that the FastMap projection method with the proposed objection function provided the best ensemble clustering results for all data sets.