SHAP Summary Plot

Features are ranked vertically by their overall impact on the model's output. The importance is calculated as the mean absolute SHAP value across all instances: mean(|SHAP value|). This provides a robust, game-theoretically grounded alternative to traditional feature importance metrics like Gini importance or permutation importance. The ranking immediately identifies which signals the model relies on most heavily, allowing for a quick sanity check against domain knowledge.

Each point on the plot represents a SHAP value for a single instance. The horizontal position shows whether the feature's impact for that instance pushed the prediction higher (positive SHAP) or lower (negative SHAP) from the baseline value. The density of points reveals the distribution of effects—tight clustering indicates a consistent linear relationship, while wide dispersion suggests complex, non-linear interactions or heterogeneous effects across the population.

Points are colored by the original feature value, typically using a diverging color scale (e.g., blue for low, red for high). This dual encoding—position on the x-axis for impact magnitude and direction, and color for the feature's actual value—allows you to instantly diagnose the directionality of a feature's effect. For example, if red points cluster on the positive SHAP side, high feature values consistently increase the prediction.

Vertical dispersion at a single feature value indicates interaction effects. If a feature has a consistent linear effect, all instances with the same feature value should have nearly identical SHAP values, forming a tight vertical band. Wide vertical spread at a given value (e.g., age=40) reveals that another interacting feature is modulating the effect. The summary plot can be colored by a secondary feature to identify the interacting variable, a pattern formally quantified by SHAP interaction values.

Extreme SHAP values appear as isolated points far from the main cluster. These represent instances where a feature had an unusually large impact on the prediction, often corresponding to outliers in the feature space or rare combinatorial patterns. Investigating these points can uncover data quality issues, edge cases where the model behaves unexpectedly, or highly influential records that merit individual audit using a SHAP waterfall plot.

For large datasets, a violin summary plot replaces individual dots with density contours, preventing overplotting while preserving the distribution shape. The standard dot plot is preferred for datasets with fewer than ~5,000 instances where individual points remain distinguishable. Both variants convey the same information: feature ranking by importance, effect directionality via color, and effect heterogeneity via horizontal spread.

The summary plot's primary function is to visualize global feature importance by aggregating local explanations. Features are ranked vertically by their mean absolute SHAP value across all instances.

• Calculation: Average of |SHAP value| for each feature over the dataset.
• Interpretation: Higher ranking indicates a larger average impact on the model's output magnitude.
• Contrast: Unlike permutation importance, this uses game-theoretic attributions that satisfy the Efficiency Property.

A SHAP dependence plot provides a granular view of a single feature's effect, which is aggregated in the summary plot. It is a scatter plot of feature value vs. SHAP value.

• Interaction Effects: Vertical dispersion at a given feature value indicates interaction with another feature.
• Color Coding: Automatically colored by an interacting feature to reveal SHAP Interaction Values.
• Use Case: Use this to drill down after identifying a high-importance feature with a wide distribution in the summary plot.

While the summary plot aggregates all instances, the SHAP waterfall plot decomposes a single prediction. It shows how each feature pushes the model output from the baseline value (expected output) to the final prediction.

• Structure: Features are displayed as horizontal bars starting from the baseline.
• Color: Red bars push the prediction higher; blue bars push it lower.
• Link: The sum of all individual contributions in the waterfall plot equals the difference between f(x) and the baseline, satisfying Local Accuracy.

An interactive visualization that stacks feature contributions to show the additive forces pushing a prediction higher or lower from the baseline value.

• Mechanism: Features that increase the prediction are shown in red, pushing right; features that decrease it are in blue, pushing left.
• Clustering: Multiple force plots can be rotated and clustered to identify distinct explanation patterns in a cohort.
• Relationship: The summary plot's color gradient (feature value) and spread (SHAP value) are the aggregated form of these individual force arrows.

The exact computation algorithm that makes summary plots for tree-based models (XGBoost, LightGBM, CatBoost) computationally feasible. TreeSHAP computes exact Shapley values in polynomial time.

• Performance: Reduces complexity from exponential O(TL2^M) to O(TLD^2) where T is trees, L is leaves, D is depth.
• Interaction Values: Directly extracts SHAP Interaction Values without approximation.
• Contrast: Unlike KernelSHAP, it does not require a background dataset for marginal expectation, using the tree structure directly.

The background dataset is the reference sample used to compute the expected model output, representing the 'missing feature' state. It directly influences the summary plot's baseline.

• Selection: Should be a representative sample (often 100-500 instances) of the training or production data.
• Impact: Using a skewed background shifts the baseline value, altering the magnitude and sign of SHAP values in the plot.
• K-Medoids: Common practice is to use k-medoids clustering to select a diverse subset that captures the data distribution.