Rotation-DPeak: Improving Density Peaks Selection for Imbalanced Data

Density Peak (DPeak) is an effective clustering algorithm. It maps arbitrary dimensional data onto a 2-dimensional space, which yields cluster centers and outliers automatically distribute on upper right and upper left corner, respectively. However, DPeak is not suitable for imbalanced data set with large difference in density, where sparse clusters are usually not identified. Hence, an improved DPeak, namely Rotation-DPeak, is proposed to overcome this drawback according to an simple idea: the higher density of a point p, the larger \(\delta \) it should have such that p can be picked as a density peak, where \(\delta \) is the distance from p to its nearest neighbor with higher density. Then, we use a quadratic curve to select points with the largest decision gap as density peaks, instead of choosing points with the largest \(\gamma \), where \(\gamma =\rho \times \delta \). Experiments shows that the proposed algorithm obtains better performance on imbalanced data set, which proves that it is promising.