Tornado forecasting with multiple Markov boundaries
Reliable tornado forecasting with a long-lead time can greatly support emergency response and is of vital importance for the economy and society. The large number of meteorological variables in spatiotemporal domains and the complex relationships among variables remain the top difficulties for a long-lead tornado forecasting. Standard data mining approaches to tackle high dimensionality are usually designed to discover a single set of features without alternating options for domain scientists to select more reliable and physical interpretable variables. In this work, we provide a new solution to use the concept of multiple Markov boundaries in local causal discovery to identify multiple sets of the precursors for tornado forecasting. Specifically, our algorithm first confines the extremely large feature spaces to a small core feature space, then it mines multiple sets of the precursors from the core feature space that may equally contribute to tornado forecasting. With the multiple sets of the precursors, we are able to report to domain scientists the predictive but practical set of precursors. An extensive empirical study is conducted on eight benchmark data sets and the historical tornado data near Oklahoma City, OK in the United States. Experimental results show that the tornado precursors we identified can help to improve the reliability of long-lead time catastrophic tornado forecasting.
