Application of ensemble methods in outlier point detection in meteorological time series

Abstract

For this research work, the performance of ensemble methods in the task of outlier points detection in meteorological univariate time series was studied, using the F1 metric to measure the performance. For this purpose, an application was created that allows applying 3 non-ensemble classifiers (support vector regression, ARIMA, bayesian networks) and 3 ensemble classifiers (stacking, bagging and AdaBoost) to 3 meteorological datasets (rainfall, maximum temperature and solar radiation).

Using this application, an experiment was executed to compare the different classifiers. In this experiment, first, the F1 average of the algorithms was obtained by executing multiple tests in each dataset. Then, using a statistical hypothesis test we compared the obtained averages to find out if the observed differences were significant. Finally, a result analysis was performed, focused on comparing the performance of the ensemble classifiers versus the performance of the best non-ensemble classifier for each dataset.

In general the results indicate that it is possible to significantly improve the performance in the outlier point detection task in some uni-variate time series by using ensemble methods. However, to obtain this improvement several conditions must be met. Although the conditions vary depending on the ensemble method, in general these conditions aim to improve the diversity in the base classifiers. When these conditions were not met, the ensemble methods didn’t have a significant difference in the performance compared to the non-ensemble classifier that got the best performance in the datasets.