Comparison of Four Machine Learning Methods for Predicting Pm10 Concentrations in Helsinki, Finland

Machine learning methods can offer a practical alternative to deterministic and statistical methods for predicting air pollution concentrations. However, for a given data set, it is often not clear beforehand which machine learning method will yield the best prediction performance. This study compares the variable selection and prediction performance of four machine-learning methods of different complexity: logistic regression, decision tree, multivariate adaptive regression splines and neural network. The methods are applied to the task of predicting the exceedance of the European PM10 daily average objective of 50 μg m −3 for a station in Helsinki, Finland. Our study shows that some predictors were selected by all models but that the different models also picked different variables. The performance of three of the four methods investigated was very similar, however, performance of the decision tree method was significantly inferior. Performance was sensitive to the learning sample size and time period used.