关键词: 时空聚类, 多时间尺度, 密度聚类, 案事件

Abstract:

Space-time clustering, which is one of the main research focuses in the field of data mining, has important application values in the field of environment protection, disease prevention and control, and crime prevention and combat. The time "distance" is considered to be a substantial interval within the existing space-time clustering methods. However, crime cases with social attributes have obvious cyclical characteristics in different time-scales. It would be difficult to find the real rules of time and space for crime cases if these characteristics are ignored. Therefore, based on DBSCAN, an algorithm considering multiple time-scales and equivalent spatio-temporal neighborhood (MTS-ESTN DBSCAN) was put forward. In this algorithm, the various time attributes in multiple time-scales were considered, the equivalent spatio-temporal neighborhood was built, and the concept of the classical density clustering algorithm was cited. In the equivalent spatio-temporal neighborhood, the Euclidean distance (L2-norm) is adopted as the measurement of spatial neighborhood for the space domain. With the improved function of HDsim, which is a method used to measure the unified similarity of high dimensional data, we defined the similarity of time domain. Based on the crime cases data in the urban area of Fuzhou city during 2013, cluster analysis was conducted, and the resultant clustering quality was evaluated using several indicators such as CH (Calinski-Harabasz), Sil (Silhouette), DB (Davies-Bouldin) and KL (Krzanowski-Lai). The results showed the feasibility of the method in space-time cluster analysis of crime cases. Compared with the traditional algorithm of ST-DBSCAN, this algorithm has produced better quality of clustering. In addition, this algorithm can find the accumulation characteristics behind the rules of human´s work, rest and other social activities in a long period. It has certain significances and application values for the advanced study of criminal geography in urban area.

Key words: space-time clustering, multiple time scale, density-based clustering, crime cases