Xgboost (classification problem) feature importance per input not for the model

Question

I have trained a xgboost model for a classification problem. I'm able to get the feature importance for the model as below.

http://machinelearningmastery.com/feature-importance-and-feature-selection-with-xgboost-in-python/

But I would like to get feature importance per input. Basically, one input may give me a output probability 0.9, another one input may give me a output probability 0.1. I want to know because which features (with values) give me a probability 0.9? which features (with values) give me a probability 0.1?

How can I approach that? Is there a package for this?

Answer 1

A partial dependence plot might be what you are after: these plots show the relationship between the probability and the input variable. The mlr package in R takes care of this.

Discussion of Partial Dependence on XGBoost Git: https://github.com/dmlc/xgboost/issues/486

General Tutorial using mlr: https://mlr-org.github.io/mlr-tutorial/devel/html/partial_dependence/index.html