SHAP waterfall: local explanation is not the same as global importance
A SHAP waterfall does not explain your whole model. It explains one prediction. That distinction changes how you use interpretability in practice.
TL;DR: Interpretability becomes much more valuable when it shifts from feature ranking to decision explanation. Explanation is not one chart for the model. It is one story per prediction.
The Problem
A SHAP summary plot can show which features tend to matter across a dataset. Useful, but that is a global view. A lot of people use SHAP summary plots and walk away saying: "now I understand the model." What they usually understand is: which features matter on average.
The business question is often local: Why did the model flag this customer? Why was this application rejected? Why did this forecast increase?
The Approach
For Customer A (predicted: CHURN):
Recharge frequency pushed prediction UP (+0.3)
Low community share pushed prediction UP (+0.2). High voice usage pushed prediction DOWN (-0.1).
For Customer B (predicted: STAY):
Recharge frequency pushed prediction DOWN (-0.15)
High data usage pushed prediction DOWN (-0.25). Low tenure pushed prediction UP (+0.1).
Same model. Same feature space. Different prediction story. That is what local SHAP values show.
Global SHAP (mean |SHAP|) helps answer: What tends to matter overall? Local SHAP (waterfall) helps answer: What drove this prediction?
Why This Matters
In business settings, the second question is often the one people actually care about. They do not ask: "What is the average importance of this variable across the validation set?" They ask: "Why did the model flag this customer?" "Why did this case get rejected?" "Why did this forecast go up?"
That is why the most useful SHAP visual is often not the summary chart. It is the local waterfall plot.
Key Takeaway
Design insight: Interpretability becomes much more valuable when it shifts from feature ranking to decision explanation. Explanation is not one chart for the model. It is one story per prediction.
Related
LightGBM vs XGBoost in 2026 →isnull() Is a Feature →Encoding Is a Modeling Decision →Categorical Encoding Cheat Sheet →Necessary Conditions vs Drivers →A Predictive Model Is Not a Decision System →Adversarial Validation →Conformal Prediction →Don't Ask Your Model to Learn What a Formula Already Knows →Categorical to Behavioral Signal →Focal Loss for Imbalanced Classification →
FAQ
What is the key takeaway from "SHAP Waterfall: Local Explanation Is Not the Same as Global Importance"?
Interpretability becomes much more valuable when it shifts from feature ranking to decision explanation. Explanation is not one chart for the model. It is one story per prediction.
Who wrote this and what is it about?
This was written by Mahmoud Trigui, Senior Data Scientist. A SHAP waterfall does not explain your whole model. It explains one prediction. Global SHAP shows what tends to matter. Local SHAP shows what drove this prediction.