All-pairwise squared distances lead to more balanced clustering

In clustering, the cost function that is commonly used involves calculating all-pairwise squared distances. In this paper, we formulate the cost function using mean squared error and show that this leads to more balanced clustering compared to centroid-based distance functions, like the sum of squared distances in $ k $-means. The clustering method has been formulated as a cut-based approach, more intuitively called Squared cut (Scut). We introduce an algorithm for the problem which is faster than the existing one based on the Stirling approximation. Our algorithm is a sequential variant of a local search algorithm. We show by experiments that the proposed approach provides better overall optimization of both mean squared error and cluster balance compared to existing methods.