This document discusses variable importances in random forests. It begins by introducing random forests and their strengths and weaknesses, specifically their loss of interpretability. It then discusses how variable importances can help recover interpretability by providing two main importance measures: mean decrease in impurity (MDI) and mean decrease in accuracy (MDA). The document focuses on MDI and presents three key results: 1) variable importances provide a three-level decomposition of information about the output, 2) importances only depend on relevant variables, and 3) most properties are lost when K > 1 in non-totally randomized trees.