SHAP Force Plot

The force plot visualizes the Efficiency Property of Shapley values in action. It begins at the baseline value (the expected model output) and sequentially stacks each feature's marginal contribution as a force vector. Features that increase the prediction push to the right (typically colored red), while features that decrease the prediction push to the left (typically colored blue). The final prediction is the exact sum of the baseline plus all individual SHAP values, demonstrating local accuracy—the explanation perfectly reconstructs the model's output for that specific instance.

Each feature in the force plot is encoded with both magnitude and directionality:

• Red arrows/segments: Features exerting positive influence, pushing the prediction higher than the baseline
• Blue arrows/segments: Features exerting negative influence, pushing the prediction lower
• Segment width: Proportional to the absolute SHAP value, indicating the strength of the feature's impact The visualization makes it immediately obvious which factors are driving a decision and in which direction, enabling rapid debugging of unexpected model behavior.

Unlike static summary plots, the SHAP force plot is designed for interactive investigation of individual predictions. Users can hover over any force segment to see the exact feature name and its precise SHAP value. For models with many features, the plot intelligently collapses smaller contributions into an 'Other features' group, preventing visual clutter while preserving the mathematical completeness of the explanation. This interactivity makes it an essential tool for model debugging and compliance auditing, where understanding a specific high-stakes decision is critical.

When applied across multiple instances, force plots can be clustered and rotated to form a cohort-level visualization. Each prediction becomes a vertical force plot, and instances are grouped by similarity of their explanation patterns. This reveals distinct decision pathways within the model—subgroups of predictions that are driven by different feature combinations. For example, in a loan approval model, one cluster might show decisions driven primarily by income and credit history, while another cluster reveals decisions dominated by debt-to-income ratio. This clustering exposes whether the model is using consistent logic across different populations.

The force plot explicitly anchors every explanation to the baseline value, which is computed as the expected prediction across the background dataset. This anchoring provides a consistent reference point, making it meaningful to compare force plots across different instances. A prediction of 0.8 from a baseline of 0.5 tells a very different story than a prediction of 0.8 from a baseline of 0.2. The visualization also supports what-if analysis: by mentally removing or negating specific force segments, practitioners can reason about how the prediction would change if a feature took a different value, approximating counterfactual reasoning directly from the explanation.

The force plot visualization is backend-agnostic, working seamlessly with both TreeSHAP (for exact Shapley values on tree-based models like XGBoost, LightGBM, and random forests) and KernelSHAP (for model-agnostic approximations on any black-box model). When powered by TreeSHAP, the force plot displays mathematically exact contributions computed in polynomial time. When powered by KernelSHAP, the plot includes contributions estimated via weighted linear regression with a Shapley kernel. The visualization itself remains identical, providing a consistent interpretability interface regardless of the underlying computation method.