Smart Swap for more efficient clustering

Local search algorithms, such as randomized and deterministic swap-based clustering, are often used for solving clustering problem. In this paper, we propose a new swap-based local search algorithm, Smart Swap, which preserves the stability of the previous solutions but is more efficient. It performs the swap by finding the nearest pair among the centroids and sorting the clusters by their distortion values. Then it swaps one of the nearest pair centroids to any position in that cluster. K-means iteration is employed to repartition the dataset and to fine-tune the swapped solution. The algorithm is easy to implement and iterates less than the previous swap-based local search algorithms. Experiments show that the proposed algorithm keeps at least 97% stability for the synthetic datasets and 0.577 of standard deviation for the real data. It is also much faster than the other swap-based algorithms.