2025-12-10 –, Thomas Paul
Tree-based machine learning models such as XGBoost, LightGBM, and CatBoost are widely used, but understanding their predictions remains challenging. SHAP (SHapley Additive exPlanations) provides feature attributions based on Shapley values, yet its assumptions — feature independence, additivity, and consistency — are often violated in practice, potentially producing misleading explanations.
This talk critically examines SHAP’s limitations in tree-based models and introduces TreeSHAP, its specialized implementation for decision trees. Rather than presenting it as perfect, we evaluate its effectiveness, highlighting where it succeeds and where explanations remain limited. Attendees will gain a practical, critical understanding of SHAP and TreeSHAP, and strategies for interpreting tree-based models responsibly.
Target audience: Data scientists, ML engineers, and analysts familiar with tree-based models.
Background: Basic understanding of feature importance and model interpretability.
Interpreting tree-based machine learning models is crucial for trust, transparency, and responsible decision-making, particularly in high-stakes or regulated domains such as finance, healthcare, and safety-critical applications. This talk is designed for data scientists, ML engineers, and analysts who regularly work with models such as XGBoost, LightGBM, and CatBoost, and want to understand how feature attribution tools perform in real-world scenarios.
The first half of the talk examines SHAP limitations in tree-based models, including instability with correlated features, interactions between variables, and sensitivity to model retraining. Participants will learn to recognize situations where SHAP attributions may be misleading, understand the impact of violating SHAP’s theoretical assumptions, and gain strategies to critically evaluate the outputs in practice.
The second half introduces TreeSHAP, the SHAP implementation optimized for decision trees. Rather than presenting it as a perfect solution, we explore how TreeSHAP can be used to test and evaluate feature importance. We discuss where it performs well, where it may still produce misleading insights, and practical considerations for applying it responsibly in production workflows. Attendees will gain actionable knowledge to interpret tree-based model predictions with rigor, skepticism, and confidence.
Yunxin holds a Bachelor’s degree in Applied Statistics from the University of Wisconsin–Madison and a Master’s degree in Applied Statistics from New York University, with a focus on data science and big data. Since completing graduate school, Yunxin has worked under Model Risk in the finance industry for the past 2.5 years, where they specialize in evaluating, validating, and interpreting complex quantitative models. Their experience spans statistical modeling, machine learning, and model risk management, with a strong emphasis on translating analytical insights into actionable business decisions.