Improved Constrained k-Means Algorithm for Clustering with Domain Knowledge

Witnessing the tremendous development of machine learning technology, emerging machine learning applications impose challenges of using domain knowledge to improve the accuracy of clustering provided that clustering suffers a compromising accuracy rate despite its advantage of fast procession. In this paper, we model domain knowledge (i.e., background knowledge or side information), respecting some applications as must-link and cannot-link sets, for the sake of collaborating with k-means for better accuracy. We first propose an algorithm for constrained k-means, considering only must-links. The key idea is to consider a set of data points constrained by the must-links as a single data point with a weight equal to the weight sum of the constrained points. Then, for clustering the data points set with cannot-link, we employ minimum-weight matching to assign the data points to the existing clusters. At last, we carried out a numerical simulation to evaluate the proposed algorithms against the UCI datasets, demonstrating that our method outperforms the previous algorithms for constrained k-means as well as the traditional k-means regarding the clustering accuracy rate although with a slightly compromised practical runtime.