Seminar by Markus Locker: Debiasing SHAP scores in tree ensembles.
Title
Debiasing SHAP scores in tree ensembles Abstract: Our research investigates a bias inherent in the SHapley Additive exPla-nations (SHAP) values used for explaining the results of tree based machine learning models. This bias leads to models being overly sensitive to features with highentropy, resulting in inflated importance scores for these features. We propose a novel method called "shrunk SHAP" to address this issue by separating the model training and prediction processes and comparing the resulting SHAP values, thus reducing the bias towards high-entropy features and providing more accurate explanations. Our algorithm is also able to enable the detection of overfitting issues at the feature level. The effectiveness of the method is demonstrated through simulations and real-world examples, highlighting the potential of "shrunk SHAP" for improving the interpretability of random forest and boosted tree models.
Online link: https://sony-research.zoom.us/j/86098451996?pwd=i0ayxO9EhkanwLSahb4uww1I... Meeting ID: 860 9845 1996 Passcode: 183496 Organized by Prof. Roberto Capobianco Reference: https://link.springer.com/article/10.1007/s10182-023-00479-7