Model Class Reliance for Random Forests

Neural Information Processing Systems 

Variable Importance (VI) has traditionally been cast as the process of estimating each variables contribution to a predictive model's overall performance.