SHAP Values

The sum of all feature attributions equals the difference between the model's prediction for a specific instance and the average baseline prediction. This property guarantees that the explanation is a complete and faithful decomposition of the output.

• Additive: f(x) = base_value + sum(SHAP values)
• Practical Use: If a predicted lead time is 14 days and the baseline is 10 days, the SHAP values for supplier reliability, distance, and seasonality will sum to exactly +4 days.

A feature that is not included in the model input—either because it is missing or was structurally excluded—is assigned an attribution of exactly zero. This prevents the explanation from hallucinating importance for absent data.

• Guarantee: SHAP(x_i) = 0 if feature i is missing
• Example: If a shipment lacks IoT temperature sensor data, the temperature feature receives zero SHAP impact, ensuring the delay explanation relies solely on available inputs like port congestion or carrier history.

If a model is retrained or modified so that a feature's contribution increases or stays the same regardless of other features, its SHAP value will never decrease. This ensures that explanations remain logically aligned with model behavior across iterations.

• Monotonicity: Improved feature impact guarantees non-decreasing attribution
• Implication: When a supplier's on-time performance becomes a stronger predictor of lead time after retraining, its SHAP importance will reflect this increase without contradiction, maintaining trust in the explanation framework.

SHAP values are derived from cooperative game theory's Shapley values, which fairly distribute a payout among players based on their marginal contributions across all possible coalitions. In machine learning, features are the players and the prediction is the payout.

• Coalition Logic: The model is evaluated with and without each feature across all possible feature subsets
• Computational Cost: Exact calculation is exponential (O(2^N)), so practical implementations like KernelSHAP and TreeSHAP use sampling or algorithmic optimizations to approximate values efficiently.

While model-specific implementations like TreeSHAP exist for gradient boosting machines and random forests, the core SHAP framework is completely model-agnostic. KernelSHAP can explain any black-box model, including LSTMs and Temporal Fusion Transformers used in lead time forecasting.

• Universal Applicability: Works with any function f(x) that produces a scalar output
• Trade-off: KernelSHAP is slower than TreeSHAP but provides the same theoretical guarantees for deep learning or ensemble models where internal structure cannot be exploited.

SHAP values operate at the local level, explaining a single prediction. However, aggregating absolute SHAP values across a dataset produces global feature importance, revealing which factors consistently drive lead time predictions.

• Local: Why is this specific shipment predicted to be 3 days late?
• Global: Across 10,000 shipments, port congestion is the dominant delay driver 40% of the time
• Visualization: SHAP summary plots combine both perspectives, showing feature importance and the direction of impact (positive/negative) for each feature.

The overarching field of methods that make complex model predictions understandable to human planners. XAI techniques like SHAP reveal the key drivers behind a predicted delay, enabling trust and auditability. Key goals include:

• Transparency: Understanding internal model mechanics
• Interpretability: Comprehending why a specific prediction was made
• Accountability: Tracing decisions to specific features for compliance

The process of using domain knowledge to create new input variables from raw logistics data to improve model accuracy. For SHAP analysis, well-engineered features are critical because they define the vocabulary of explanation. Examples include:

• Rolling averages of transit times per carrier
• Supplier lead time variance over trailing 90 days
• Port congestion indices derived from AIS vessel data
• Seasonal decomposition of historical demand patterns

The phenomenon where the statistical properties of the target variable change over time, degrading model accuracy. SHAP values help detect drift by revealing when feature contributions shift unexpectedly. For instance, if a carrier's historical reliability score suddenly becomes the dominant negative contributor to predictions, it signals a systemic change in that carrier's performance requiring investigation.

Statistical methods that identify the root cause of supply chain failures, moving beyond correlation to causation. While SHAP values quantify feature importance, they do not establish causality. Integrating SHAP with causal frameworks like do-calculus or structural causal models allows planners to answer counterfactual questions: 'Would the delay have occurred if we had used a different carrier?'

The continuous tracking of deployed model performance and input data distributions to detect degradation. SHAP value distributions are a key monitoring metric. A monitoring dashboard typically tracks:

• Prediction drift: Changes in model output distribution
• Feature drift: Shifts in input variable distributions
• SHAP drift: Changes in feature contribution patterns over time
• Performance decay: Increasing error metrics like MAPE

An analytical capability allowing planners to alter input variables and instantly simulate the cascading impact on predictions. SHAP values power these simulations by quantifying the marginal effect of changing a single feature. For example, a planner can ask: 'What happens to the delivery date if we switch from ocean freight to air freight?' and see the predicted lead time shift based on the SHAP contribution of the transport mode feature.