I have trained an XGBoost model for prediction. The algorithm is able to calculate variable importances. I was asked why I have not analyzed these variable importances and I did not because as I understood XGBoost is rather a predictive than a descriptive model. I am not sure what extent theses importances can be used to capture real importance (I understand that it may be dependent on the context). Am I right that XGBoost is rather a predictive model?
You're right that it is a predictive model, but analysing feature importance, partial dependence and other metrics/plots are intended to allow you to gain an understanding of the classification process of models which are often very black-box in nature.
Yes, you can use this to gain some understanding of the underlying data, but I'd be more inclined to say that the primary reason for looking at feature importance is to understand your model rather than to understand your data.