Instance-based Explanations for Gradient Boosting Machine Predictions with AXIL Weights
This work addresses the need for interpretability in machine learning, particularly for complex models like Gradient Boosting Machines, by offering a novel explanation method, though it appears incremental as it builds on existing additive explanation frameworks.
The authors tackled the problem of explaining predictions from tree-based ensemble models by introducing AXIL weights, which represent predictions as linear combinations of training instances, providing both local and global explanations. The result is a new measure of instance importance that complements existing feature importance methods like SHAP and LIME.
We show that regression predictions from linear and tree-based models can be represented as linear combinations of target instances in the training data. This also holds for models constructed as ensembles of trees, including Random Forests and Gradient Boosting Machines. The weights used in these linear combinations are measures of instance importance, complementing existing measures of feature importance, such as SHAP and LIME. We refer to these measures as AXIL weights (Additive eXplanations with Instance Loadings). Since AXIL weights are additive across instances, they offer both local and global explanations. Our work contributes to the broader effort to make machine learning predictions more interpretable and explainable.