Dynamic k-means clustering for risk decision making and its application to China's customs targeting

• Authors: Yu Wang;Hongshan Xiao
• Venue: Proceedings of the 14th Annual International Conference on Electronic Commerce
• Year: 2012

Customs targeting is a typical risk decision making problem. In this problem, the empirical smuggling probability density function of import/export goods is needed for targeting decision. Generally, the density function can be obtained by applying statistical analysis, especially regression analysis, to historical observations (samples). A critical presumption is that the samples are homogeneous, which means they are drawn from the same distribution. Therefore, clustering techniques are usually employed as the preprocessing methods in statistical analysis. However, in China's customs targeting problem, severe heterogeneity and abnormality exist in the historical observations, which makes the conventional clustering methods inapplicable for preprocessing. In this paper, we develop a dynamic K-means clustering approach to solve this problem. Through optimizing a validity function of clustering, the proposed approach divides the entire samples into different clusters iteratively. As the result of this preprocessing technique, samples in the same cluster are more compact, while in different clusters are more discriminated. Based on the proposed dynamic clustering approach, we develop a risk decision making process. Application to China's customs targeting problem indicates that the proposed approach could increase the efficiency of customs targeting decision.