[4Rin1] Interactive 2

[4Rin1-26] Variable Importance Measures Based on Occurrence Ratio for Tree Ensembles

〇Kota Mata¹, Kentaro Kanamori¹, Hiroki Arimura¹ (1.Hokkaido University)

Keywords:tree ensembles, feature importance, interaction term

As one of the methods for interpreting the prediction result of a complex machine learning model, variable importance scores of a tree ensemble are widely used. However, existing variable importance methods measure how much a single variable has contributed to the output. Therefore, when a relationship between two or more variables (such as XOR) contributes to the output, their importance scores are often not accurately evaluated. In this paper, we propose a new variable importance score based on the occurrence ratio of a variable in a tree ensemble. By experiments, we confirm the effectiveness of our proposed method in comparison with existing methods.