A Robust Seedless Algorithm for Correlation Clustering

Finding correlation clusters in the arbitrary subspaces of highdimensional data is an important and a challenging research problem. The current state-of-the-art correlation clustering approaches are sensitive to the initial set of seeds chosen and do not yield the optimal result in the presence of noise. To avoid these problems, we propose RObust SEedless Correlation Clustering (ROSECC) algorithm that does not require the selection of the initial set of seeds. Our approach incrementally partitions the data in each iteration and applies PCA to each partition independently. ROSECC does not assume the dimensionality of the cluster beforehand and automatically determines the appropriate dimensionality (and the corresponding subspaces) of the correlation cluster. Experimental results on both synthetic and real-world datasets demonstrate the effectiveness of the proposed method. We also show the robustness of our method in the presence of a significant noise levels in the data.