Data Mining to Classify Fog Events by Applying Cost-Sensitive Classifier

The present work illustrates a Data Mining application in the meteorological domain. In particular, this work illustrates the creation of some fog classifying local indices, based on the post-processing of meteorological variables. A dataset containing a total amount of 17396 records, collected in Trapani Milo station and a poor quote of 142 fog events was obtained, such rare event required some specific approaches of the Data Mining techniques in order to overcome the class imbalance problem: a Cost Sensitive Classifier (matched with Bayes Network). The obtained results were evaluated by means of adequate performance metrics able to highlight the classifying ability of an index with respect to the fog events and the no-fog events separately (confusion matrix, ROC, AUC, ...). The obtained models were tested over 4349 records; four models overcame the AUC threshold of 0.8 and, for one of them, the ROC curve showed a good result: 88% of fog events correctly predicted.