XGBoost: feature importance changes after bigger training set

Question

I predict PM10 values in the air with XGBoost model. My main concern is feature importance. I have two types of experiments: one on the longer history dataset (training size: 4