Comparison of Four Machine Learning Methods for Predicting PM10 Concentrations in Helsinki, Finland

Mindaugas Zickus

Abstract: Machine learning methods can offer a practicalalternative to deterministic and statistical methods forpredicting air pollution concentrations. However, for agiven data set, it is often not clear beforehand whichmachine learning method will yield the best predictionperformance. This study compares the variable selection andprediction performance of four machine-learning methods ofdifferent complexity: logistic regression, decision tree,multivariate adaptive regression splines and neuralnetwork. The methods are applied to the task of predictingthe exceedance of the European PM10 daily averageobjective of 50 μg m-3 for a station in Helsinki,Finland. Our study shows that some predictors were selectedby all models but that the different models also pickeddifferent variables. The performance of three of the fourmethods investigated was very similar, however, performanceof the decision tree method was significantly inferior.Performance was sensitive to the learning sample size andtime period used.