The invention discloses an improved LR-Bagging algorithm based on characteristic selection. The improved LR-Bagging algorithm comprises the following steps: first of all, determining an initial data set from original data, wherein a degree of correlation between independent variables and dependent variables is required not to be excessively low; next, carrying out WEO encoding on discrete independent variables in the initial data set; then, obtaining a certain number of records and characteristic fields by employing random sampling to form training examples; carrying out LR (Logistic Regression) model training on the training examples and carrying out normal significance testing on coefficients; if the coefficients are not significant, eliminating the coefficients; or otherwise, adding the coefficients to a combined model; carrying out loop iteration till the combined model is relatively optimal; and at last, carrying out prediction and grouping by adopting the relatively optimal combined model. According to the algorithm, the diversity of classification results, the extracting degree of variable information and the accuracy rate of a prediction result can be improved; and the possibilities of multicollinearity and ''overfitting'' due to excessive variables of the LR model also can be effectively reduced.