Density peaks clustering with spatiotemporal nearest neighbors

Abstract

Density peaks clustering (DPC) encounters three limitations when applied to spatiotemporal data. Its local density estimate has limited capacity to distinguish samples drawn from regions with different densities, which can bias cluster center identification. Its distance measure emphasizes spatial information and may fail to separate clusters that are spatially close but temporally distinct. Its assignment strategy also lacks an explicit mechanism for anomalous samples, thereby reducing clustering accuracy. To address these limitations, a method termed density peaks clustering with spatiotemporal nearest neighbors (STN-DPC) is proposed. Separate sets of spatial and temporal K-nearest neighbors are constructed, and natural nearest neighbors are identified through mutual inclusion. Spatial and temporal base densities are then combined, while neighborhood support is weighted according to the number of natural nearest neighbors. This design improves density discrimination across heterogeneous local structures. Next, spatial and temporal average nearest neighbor distances are derived from nearest neighbor order. They correct distance shifts caused by scale differences between dimensions and limit the dominance of either dimension during cluster discrimination. Finally, a joint score is formed from the medians and median absolute deviations of local density and relative distance. Anomalous samples are identified from this score and removed from the clustering result. Experiments on the four synthetic datasets show that STN-DPC performs well in terms of AMI, ARI, and FMI, and its application to the earthquake catalogue identifies three sequences with spatiotemporal associations.