I have a supervised binary classification problem. I tuned an xgboost model on the training set and achieved a reasonably high accuracy on the test set. Now I want to interpret the results of the model.
I used the SHAP library to interpret the results and for the most part, they are consistent with what I would expect. However, there is one feature that, if we average over all shap values, is ranked as 7th most important and I would have expected it to have been higher. If I perform a t-test of the feature between the positive group and the negative group, there is a clear statistical difference between the two (p<<0.05) which implies that the feature should be very predictive of the class. What could be the cause of this discrepancy?