A new projection-based K-Means initialization algorithm

As a partition based clustering algorithm, K-Means is widely used in many areas for the features of its efficiency and easily understood. However, it is well known that the K-Means algorithm may get suboptimal solutions, depending on the choice of the initial cluster centers. In this paper, we propose a projection-based K-Means initialization algorithm. The proposed algorithm first employ conventional Gaussian kernel density estimation method to find the highly density data areas in one dimension. Then the projection step is to iteratively use density estimation from the lower variance dimensions to the higher variance ones until all the dimensions are computed. Experiments on actual datasets show that our method can get similar results compared with other conventional methods with fewer computation tasks.