(A-33) Model-Dependent Characteristics of Model-Agnostic Variable Importance Analysis in Machine Learning

Thursday, October 12, 2023

9:30 AM – 10:30 AM PDT

Location: Exhibit Hall - Row A - Poster # 33

Presenting Author(s)

HJ

Hyun Jeong

Student
Fordham University
Potomac, Maryland, United States

Primary Investigator(s)

IK

Ibrahim Karabayir

Academic Faculty Physician
Wake Forest School of Medicine, United States

Last Author(s)

OA

Oguz Akbilgic

Academic Faculty
Wake Forest School of Medicine, United States

Introduction:: Variable importance methods have emerged as a crucial area of study in the field of biomedical informatics, primarily driven by the necessity to provide more transparent explanations for the decision-making process of machine learning algorithms. Model-agnostic methods have gained significant attention due to their assertion of independence from the specific machine learning model and the training dataset utilized. The primary objective of this research was to assess the reliability and effectiveness of model-agnostic variable importance methods, particularly focusing on the SHAP (SHapley Additive exPlanations) technique, by examining its dependency on the training dataset employed during the construction of the machine learning model.

Materials and Methods:: We obtained three open-source datasets from the biomedicine field and stratified each dataset into two subsets: a cross-validation set, used as the training data to build the model, and a holdout set, used as a test set to evaluate the model’s performance. We trained three different algorithms—extreme gradient boosting, light gradient-boosting machine, and logistic regression—two times each (a well fitted and overfitted model by changing model hyperparameters). We obtained the ROC AUC, as well as the model-specific and model-agnostic feature importance for both the train and test set. Further, we manipulated the test set by replacing several key variables with a single random value, then performed the same analysis on this modified test set. We employed the Kruskal-Wallis and Friedman tests to quantitatively compare the relative variable importances across the four experimental setups: model-specific, model-agnostic on train set, test set, and modified test set. We employed the Python implementations of XGBoost, LightGBM, Logistic Regression, and SHAP libraries.

Results, Conclusions, and Discussions:: Our findings show that regardless of whether the models were overfitted or not, the model-specific and model-agnostic variable importances are consistent. The Friedman test results, outlined in Table I, show that there is no significant variance between the relative variable importance values across the four experimental conditions at a 99% confidence interval. Figure 1, outlining the relative feature importance values for the overfitted LGBM model, indicates that the top four features maintained their ranking between the model-specific and model-agnostic approaches. Further, the SHAP values for the test and modified test sets showed minimal change in the top five variables, suggesting that the stark difference induced between the evaluation data and the model accuracy had little influence to feature importance.

We were able to quantitatively evaluate the similarities in variable importance rankings across various machine learning algorithms and datasets. We found that model-agnostic methods like SHAP are not sensitive to model performance nor the manipulations of the test set. We showed that the relative variable importance of certain machine learning models depend entirely on the specific model and the data used to train that model. The results highlight the importance of including training and evaluation set accuracy metrics in studies utilizing model-agnostic methods. Since variable importance analysis heavily relies on the characteristics of the training dataset, the inclusion of such metrics is crucial for the robustness, interpretability, and reliability of the variable importance analysis. Future studies could focus on developing an improved model-agnostic variable importance metric that dynamically incorporates the model performance (accuracy, ROC AUC) on the unseen test set as well as the training data.

Acknowledgements (Optional): :

This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Biomedical Informatics at Wake Forest University School of Medicine.

References (Optional): : [1] Janosi, Andras, Steinbrunn, William, Pfisterer, Matthias, and Detrano, Robert. (1988). Heart Disease. UCI Machine Learning Repository. https://doi.org/10.24432/C52P4X.
[2] Little, Max. (2008). Parkinsons. UCI Machine Learning Repository. https://doi.org/10.24432/C59C74.
[3] Wolberg, William, Mangasarian, Olvi, Street, Nick, and Street, W. (1995). Breast Cancer Wisconsin (Diagnostic). UCI Machine Learning Repository. https://doi.org/10.24432/C5DW2B.