Subspace Similarity-Based Algorithm for Combine Multiple Clustering

Ensemble learning methods train multiple classifiers before classification combination. The methods have been proved to be very effective in supervised machine learning. In this paper, we present an approach to solve ensemble problem of clustering. Beginning with pursuing a "best" subspace, we formulate the problem as an optimization of square sum of Euclidean distances between the standard orthogonal basis of the target subspace and the given subspace sets. We then reach the status that the low dimensional embedding of instances and hyper-edges are simultaneously attained. Finally, we use the K-mean algorithm in optimization principle to cluster instances according to their coordinates in the embedding space. This way, we obtain stable clustering results. We apply our ensemble algorithm on several well-recognized datasets. After comparing our experimental results with others, can conclude that our algorithm outperforms other algorithms in terms of the normalized mutual information.