Inspection data of actual concrete structures should be analyzed to elucidate the deterioration mechanism and construct a regression model. Although machine learning can be applied to this problem, inspection data are not suitable because machine learning targets big data with a uniform density and a balanced distribution. This study applies machine learning to a regression model of the crack damage grade in concrete bridges, using imbalanced inspection data. The model performance is improved by analyzing the influence of undersampling. Undersampling is conducted step-wise, and the models are constructed by learning all the undersampled data. The cross-validation of these models yielded the regression errors on each crack damage grade to evaluate the model performance considering the bias of data imbalance. Based on the results, the effect of undersampling on the model performance is analyzed, and the appropriate model is selected. Additionally, the influence of the model difference on the evaluation is investigated via historical change or factor analysis to confirm the effect of undersampling. This article not only presents a case study of a regression task for crack damage grades in concrete bridges, but also describes a strategy to maximize the use of imbalanced data for regression problems.