Model Selection of Symbolic Regression to Improve the Accuracy of PM2.5 Concentration Prediction

As one of the main components of haze, topics with respect to PM2.5 are coming into people’s sight recently in China. In this paper, we try to predict PM2.5 concentrations in Dalian, China via symbolic regression (SR) based on genetic programming (GP). During predicting, the key problem is how to select accurate models by proper interestingness measures. In addition to the commonly used measures, such as R-squared value, mean squared error, number of parameters, etc., we also study the effectiveness of a set of potentially useful measures, such as AIC, BIC, HQC, AICc and EDC. Besides, a new interestingness measure, namely Interestingness Elasticity (IE), is proposed in this paper. From the experimental results, we find that the new measure gains the best performance on selecting candidate models and shows promising extrapolative capability.