When using surrogate-assisted evolutionary algorithm to solve the expensive many-objective optimization problems, the surrogate is usually used to approximate the expensive fitness function. However, with the increase of the number of objectives, the approximation error will accumulate gradually and the amount of calculation will increase sharply. In order to solve this problem, we propose an improved ensemble learning classification based surrogate-assisted evolutionary algorithm, which uses an improved bagging ensemble as the surrogate. Firstly, a set of classification boundary individuals are selected from the individuals evaluated by the expensive fitness function, and the individuals are divided into two groups. Then, these individuals with the group labels are used to train a classifier to predict the groups of the candidate individuals. Finally, the promising individuals are selected to be evaluated by the expensive fitness function. The experimental results show that the proposed surrogate in the algorithm effectively improves the ability of the classification based surrogate-assisted evolutionary algorithm to solve the expensive many-objective optimization problems, and compared with the current popular surrogate-assisted evolutionary algorithms, the proposed improved ensemble learning classification based surrogate-assisted evolutionary algorithm is more competitive.