基于梯度提升树的ALE图特征解释效果分析

Effect Analysis of ALE Plots Feature Interpretation Based on Gradient Boosted Trees

理解特征效应有助于机器学习模型进行调优与预测结果解读,累积局部效应(ALE)图是一种独立于模型的黑盒预测特征解释新方法,文章基于梯度提升树分析其稳定性与解释效果。首先,推导输入特征服从联合高斯分布的累积局部效应与偏相关函数,分析特征间相关性的影响;其次,对估计的稳定性与特征效应解释的效果进行模拟研究,并提出将纵坐标统一设置用于特征选择参考;最后,针对实际数据集给出ALE图的应用实例,并将特征解释结论与传统统计模型显著性分析进行对比。

Understanding feature effects helps machine learning models to debug and interpret prediction results.Accumulated local effects(ALE)plot is a new model-agnostic interpretation method for feature interpretation of black-box prediction models.This paper analyzes its stability and interpretation effect based on gradient boosted tree.First,accumulated local effects and partial correlation function are derived for the dataset whose input features follow joint Gaussian distribution.This paper is based on gradient boosted trees to analyze its stability and interpretation effect.The paper first derives the cumulative local effect and partial correlation function of input features obeying joint Gausian distribution,analyzing the influence of correlation between features,and then simulates the stability of the estimation and the effect of the feature effect interpretation,also proposing a unified ordinate setting for the reference of feature selection.Finally,the paper presents an application example of ALE plots based on the actual dataset,and compares the conclusion of feature interpretation with the significance analysis of traditional statistical model.