Abstract:

After reviewing current researches on early warning, it is found that "bad" data of some systems is not easy to obtain, which makes methods proposed by these researches unsuitable for monitored systems. An interactive early warning technique based on SVDD (support vector data description) is proposed to adopt "good" data as samples to overcome the diculty in obtaining the \bad" data. The process consists of two parts: (1) A hypersphere is fitted on \good" data using SVDD. If the data object are outside the hypersphere, it would be taken as "suspicious"; (2) A group of experts would decide whether the suspicious data is "bad" or "good", early warning messages would be issued according to the decisions. And the detailed process of implementation is proposed. At last, an experiment based on data of a macroeconomic system is conducted to verify the proposed technique.