Mixed Integer Linear Programming (MILP) is a type of NP-hard problem that is widely used in industries but challenging to solve. Recently, there has been increasing interest in using Deep Learning (DL) to accelerate solving MILPs, particularly through learning to branch. However, the validity of DL models is based on the assumption that the test set and the training set are drawn from the same distribution, while in MILPs, the distribution shift often appears. To improve the generalization of the branching policy, we introduce the domain generalization method and propose an Adaptive Policy Switching (APS) framework. The experiments demonstrate that APS outperforms the previous state-of-the-art (SOTA) method, reducing the number of nodes by 18% on completely unseen MILPs.