Stability of SHAP Explanations for Tree Models under Data Perturbation: A Multi-Model Empirical Study Based on AI4I 2020

Stability of SHAP Explanations for Tree Models under Data Perturbation: A Multi-Model Empirical Study Based on AI4I 2020

Explainable AI and SHAP Explanation Stability in Predictive Maintenance

2026-10-03

Bo Zhou

Research has found that SHAP's interpretation of model predictions varies with data Sometimes, the predicted results are similar, but the model's interpretation has indeed changed. This research project will study the patterns of change .This project uses AI4I 2020 predictive maintenance data as the dataset, which is divided into 70%, 15%, and 15% fixed partitions. After preprocessing, XGBoost and Random Forest are used to make predictions and use TreeSHAP to interpret the prediction results. In this project, four different degrees of interference will be added to five types of features, and 400 basic test samples, 3 model seeds, 110400 pairs of original perturbation samples, and 883200 feature attribution records will be analyzed In the stable dataset predicted in this experiment, as Gaussian noise increased from 0.01 to 0.20, the explanatory change score of XGBoost increased from 0.047 to 0.201, and the explanatory change score of Random Forest increased from 0.008 to 0.143. The overlap rate of Top-5 was 1.0. In addition, among the 90582 events with a failure probability change rate not exceeding 0.02, there were only two strictly unstable events, both of which were feature exchange positions with similar SHAP values. The difference in interpretation will increase with the increase of data. If you only look at the top value, you may not be able to see subtle differences .Accurate model prediction does not mean stable explanation; Explaining stability does not necessarily mean that the explanation is correct, nor does it mean that it proves causal relationships in reality.

Explainable AI, SHAP, Machine Learning, Predictive Maintenance, Model Interpretability, Explanation Stability, XGBoost, Random Forest