Mining Multiple Temporal Patterns of complex dynamic data systems

We present a new data mining method, called Multiple Temporal Pattern Recognition (MTPR), that is capable of mining and detecting multiple temporal patterns for characterizing and predicting significant events in the complex dynamic system data. The MTPR method first embeds the time series data into multiple phase spaces with various dimensions and time delays. Then it clusters the embedded data to identify the preliminary temporal patterns. The new method further performed a three-stage predictability analysis to evaluate the preliminary temporal patterns and detect those with high confidence. This is accomplished by first introducing a new Predictability Measure, pm, to evaluate the effectiveness of the detected temporal patterns and then apply the statistical logistical regression to further validate these patterns. Experimental results demonstrated effectiveness the proposed MTPR method, especially in the complex time series setting.