Improving atmospheric corrosion prediction through key environmental factor identification by random forest-based model

Abstract Atmospheric corrosion prediction models were constructed based on the corrosion rates of carbon steel and 12 environmental factors from long-term exposure tests. Prior to support vector regression (SVR) modelling, the dimensionality of the dataset was reduced by a hybrid method combining random forest (RF) and Spearman correlation analyses, compared with maximal information coefficient (MIC) and principal component analysis (PCA). Using key environmental factors identified by the hybrid method as input parameters, the SVR model presented higher accuracy than those with dimensionality reduction by MIC and PCA. The dimensionality reduction also significantly improved the accuracy and generalizability of the SVR model.