Attribute weights-based clustering centres algorithm for initialising K-modes clustering

The K-modes algorithm based on partitional clustering technology is a very popular and effective clustering method; moreover, it handles categorical data. However, the performance of the K-modes method is largely affected by the initial clustering centres. Random selection of the initial clustering centres commonly leads to non-repeatable clustering result. Hence, suitable choice of the initial clustering centres is crucial to realizing high-performance K-modes clustering. The present article develops an initialisation algorithm for K-modes. At initialisation, the distance between two instances calculated after weighting the attributes of the instances. Many studies have shown that if clustering is based only on distances or density between the instances, the clustering revolves around one centre or the outliers. Therefore, based on the attribute weights, we combine the distance and density measures to select the clustering centres. In experiments on several UCI machine learning repository benchmark datasets, the new initialisation method outperformed the existing K-modes clustering methods.