# How to interpret feature importance (XGBoost) in this case?

I found two dominant features from plot_importance. My dependent variable Y is customer retention (whether or not the customer will retain, 1=yes, 0=no). My problem is I know that feature A and B are significant, but I don't know how to interpret and report them in words because I can't tell if they have a position or negative effect on the customer retention. Is there a way to find that out or anything that helps make it clear?

Thanks.

## 3 Answers

Pictures usually tell a better story than words - have you considered using graphs to explain the effect?

Perhaps 2-way box plots or 2-way histogram/density plots of Feature A v Y and Feature B v Y might work well.

I think you can find the correlation matrix for the feature which could provide you with evidence to justify your hypothesis.

• The target - Y - is binary. Correlation measures the relationship between two continuous features and so is inappropriate to use in this case. – bradS Jun 21 at 8:30

Put it in another way. If we would not know this information we would be %point less accurate. Calculate accuracy using your model, then shuffle your variable to explain randomly, predict with your model and calculate accuracy again. The difference will be the added value of your variable. (its called permutation importance)

If you want to show it visually check out partial dependence plots. (read more here)

It is also powerful to select some typical customer and show how each feature affected their score. (i.e. using SHAP values see it here)