Shift of pairwise similarities for data clustering

Abstract Several clustering methods (e.g., Normalized Cut and Ratio ) divide the Min cost function by a cluster dependent factor size or degree of clusters), in order to yield more balanced partitioning. We, instead, investigate adding such regularizations original function. We first consider case where regularization term is sum squared clusters, then generalize it adaptive pairwise similarities. This leads shifting (adaptively) similarities which might make some them negative. study connection this method Correlation Clustering propose an efficient local search optimization algorithm with fast theoretical convergence rate solve new problem. In following, we shift on common methods, finally, demonstrate superior performance extensive experiments different datasets.