Most of the existing data-driven surrogate-assisted optimization algorithms are developed to solve continuous optimization problems, but there are many problems in reality belong to combinatorial optimization/discrete optimization problems, especially (subdivided into) mixed-integer optimization problems. However, there are few researches based on online data-driven mixed-integer linear programming. Therefore, in this work, we solve a kind of expensive data-driven constrained multi-objective mixed-integer optimization problem, the objectives and constraints are derived based on a large number of calculations. In order to solve this kind of problem, we proposed the use of a random forest classifier (RF) as a surrogate model to approximate the objective and constraints. At the same time, to balance the convergence, diversity, and complexity of the objective, this paper proposes a new improvement strategy that combines the surrogate-assisted model and the Two_Arch method, which assigns different selection principles to the two files, and a dual structure multi-objective maintenance program based on Hypervolume district size (HD) indicators. To further improve the accuracy and performance of the model, this paper also adopts a new crossover operator. To verify the effectiveness of the algorithm, tests are carried out on ten benchmark problems of multi-objective knapsack, and comprehensive comparison with several well-known algorithms is carried out. Experimental results show that the improved algorithm is effective in solving data-driven multi-objective mixed-integer optimization problems.