An Improved Random Forest Model Combined with Bootstrap and Under sampling for Urban Management Case Classiftcation

Digital grid management is a promising direction of urban management, while the current classification of urban management cases in the digital urban management system is mainly based on the subjective judgment, which might result in inefficiency and even errors. In classification, the random forest (RF) is an efficient and powerful algorithm, yet it performs less satisfactorily due to its traditional Bootstrap resampling when faced with a large amount of imbalanced data. In this study, we propose an improved random forest model combined with Bootstrap and Under sampling (RFCBU) to address this problem. In RFCBU, we employ both Bootstrap and Under sampling to train some of the classification and regression trees (CART). To evaluate its validity, we compare it with 4 basic models, namely logistic regression (LR), gradient boost decision tree (GBDT), extreme gradient boosting (XGB), and the original random forest (RF), with the data from Shenzhen digital management system. Experimental results verify the validity of RFCBU in terms of accuracy and receiver operating curve (ROC).