Forest Fires Dataset

In [Cortez and Morais, 2007], the output 'area' was initially transformed using a ln(x+1) function. Several Data Mining techniques were then applied to the data. After fitting the models, the outputs were post-processed with the inverse of the ln(x+1) transformation. Four different input configurations were utilized for the experiments, which were carried out using a 10-fold cross-validation with 30 runs. Two regression metrics, MAD and RMSE, were measured. A Gaussian support vector machine (SVM) using just 4 direct weather conditions (temperature, relative humidity, wind, and rain) achieved the best MAD value: 12.71 ± 0.01 (mean and confidence interval within 95% using a t-student distribution). The naive mean predictor attained the best RMSE. An analysis of the regression error curve (REC) indicated that the SVM model performed better in predicting more examples with a lower admitted error. Specifically, the SVM model was more effective at predicting smaller fires, which constitute the majority of the cases.