This recursive binary splitting naturally segments the feature space in a way that can capture non-linear relationships. For instance, imagine a U-shaped relationship between a feature and the target variable. A decision tree could make one split to isolate the left arm of the U and another split to isolate the right arm, thereby approximating the non-linear shape. NN has a lot of problems with tabular data. An NN to create boundaries like decision trees often needs a lot of data; it takes time for the NN to settle its decision boundaries with sufficient data. Also NN doesn't handle missing data properly thus xgboost outperforms NN.