K-means clustering based on self-adaptive weight
Yuzhu Zhang, Damin Zhang, Hualin Shi · 2012
Traditional K-means algorithm is sensitive to the initial cluster centers. It is easily influenced by the outlier and the uneven distribution of the samples to fall into the local optimum states. The K-means algorithm treats all features fairly and sets weights of all features equally when evaluating dissimilarity, which easily makes the cluster fall into the trap of dimension. Due to the deficiency of k-means algorithm, this paper proposed an improved K-means clustering algorithm based on self-adaptive weights. The new method uses the Gaussian distance ration in cluster approximation to set the sample weights. It computed the weights for every data according to the current clustering state and did not rely on the initial clustering center and initial weights any more. The new method selects the weights more scientifically and gets more accurate clustering results. The experimental results show that the algorithm accuracy and stability are significantly higher than the traditional K-means algorithm.