SHAP Values

The sum of all feature attributions plus the base value (average model output) always equals the model's actual prediction for that specific instance. This property guarantees that the explanation is a complete and faithful decomposition of the output, leaving no residual variance unaccounted for.

• Mathematical Guarantee: f(x) = φ₀ + Σᵢ φᵢ
• Practical Meaning: You can literally add up the SHAP values to reconstruct the prediction score.
• Contrast: Unlike LIME, which fits a local surrogate that may not perfectly match the original prediction, SHAP is an exact additive decomposition.

A feature that is not present in the input (i.e., its value is missing or it was not used by the model) is guaranteed to receive a SHAP value of zero. This prevents the explanation from arbitrarily assigning importance to features that did not influence the decision.

• Zero Attribution: If xᵢ is absent, φᵢ = 0.
• Structural Integrity: Ensures that the explanation respects the actual feature space used by the model.
• Implementation Note: This is handled by manipulating the feature's value to a background dataset reference during the estimation process.

If a model is changed so that a feature's contribution increases or stays the same regardless of other inputs, the SHAP value for that feature cannot decrease. This ensures that explanations are logically stable and that a feature's attributed importance monotonically tracks its true impact.

• Monotonicity: If a feature becomes more important to the model, its SHAP value must reflect that.
• Model Comparison: This property is critical for comparing two different models and trusting that a higher SHAP value genuinely means a larger effect.
• Violation Example: Gain-based feature importance in tree models can violate this, making SHAP a more reliable alternative.

SHAP is a unified framework that connects game theory with local explanations. While it has model-specific fast implementations (TreeSHAP, DeepSHAP, LinearSHAP), the core KernelSHAP algorithm is fully model-agnostic and can explain any black-box function.

• KernelSHAP: Uses a weighted linear regression on coalitions of features to estimate Shapley values for any model.
• TreeSHAP: An exact, polynomial-time algorithm for tree-based models (XGBoost, LightGBM, Random Forest) that computes SHAP values without sampling error.
• DeepSHAP: A high-speed approximation for deep learning models that combines SHAP with DeepLIFT.

Beyond individual feature effects, SHAP can decompose predictions into Shapley interaction values. These capture the synergistic or antagonistic effect between pairs of features that cannot be explained by their individual contributions alone.

• Pairwise Effects: φᵢⱼ represents the pure interaction effect between feature i and feature j.
• Fraud Use Case: In transaction monitoring, an interaction between a high transaction amount and a sudden change in device fingerprint might have a disproportionately large effect that individual SHAP values miss.
• Visualization: SHAP dependence plots can be colored by an interacting feature to visually reveal these non-linear synergies.

While SHAP values are locally computed for a single prediction, aggregating them across a dataset provides powerful global insights. A SHAP summary plot combines feature importance and feature effects into a single visualization.

• Mean |SHAP|: The average absolute SHAP value for a feature indicates its global importance.
• Feature Density: The spread of SHAP values reveals how a feature's impact varies across the population.
• Directionality: The sign of the SHAP value shows whether a feature pushes the prediction higher or lower, distinguishing it from purely magnitude-based importance metrics.

SHAP borrows from cooperative game theory to assign a fair payout to each feature for a prediction. It calculates the marginal contribution of a feature by comparing the model's output with and without that feature, averaged over all possible subsets of features. This ensures the sum of all feature attributions equals the difference between the actual prediction and the average prediction. For a fraud model, this means the exact dollar amount of a transaction's risk score can be decomposed and assigned to specific drivers like transaction amount, device fingerprint mismatch, or velocity of recent logins.

SHAP operates on two critical levels for fraud analysis:

• Local Explanations: Explain a single prediction. Why was this specific transaction flagged? A force plot can show that a high cross-border transfer amount pushed the risk score up by 0.4, while a recognized device ID pushed it down by 0.1.
• Global Explanations: Aggregate local explanations to understand overall model behavior. A summary plot reveals that transaction amount is the most impactful feature globally, but time since last password change has a long tail of high-impact outliers, indicating its role in catching account takeover fraud.

Unlike permutation importance, which measures a feature's overall impact on model performance, or built-in Gini importance from tree models, SHAP provides consistent, individualized explanations. Key advantages:

• Consistency: If a model relies more on a feature, its SHAP value magnitude will not decrease.
• Accuracy: SHAP respects feature interactions. Two features like transaction amount and account age might have a synergistic effect that simpler additive methods miss.
• Model-Agnostic: KernelSHAP works with any model, while TreeSHAP is optimized for tree-based ensembles like XGBoost and LightGBM, which are common in fraud detection.

Financial regulators increasingly require adverse action reasons for automated credit and fraud decisions. SHAP values provide a mathematically sound basis for generating these reasons. For a declined transaction, the explanation can be automatically generated: 'This transaction was blocked due to an unusually high transaction amount relative to the account's 30-day average and a device location inconsistent with the cardholder's home address.' This bridges the gap between a black-box deep learning model and the right to explanation mandated by regulations like GDPR and the Fair Credit Reporting Act (FCRA).

Global SHAP analysis is a powerful debugging tool. If a protected attribute like geographic region or a proxy for it shows unexpectedly high SHAP importance, it signals potential bias requiring investigation. Similarly, if a feature like transaction ID or a highly granular user agent string has a dominant SHAP value, it strongly indicates data leakage—the model is memorizing identifiers rather than learning generalizable fraud patterns. This allows MLOps teams to catch critical flaws before a model is deployed to production.

Calculating exact Shapley values is computationally exponential. However, TreeSHAP, a variant specific to tree-based models, reduces the complexity to O(TLD^2) where T is the number of trees, L is the maximum number of leaves, and D is the maximum tree depth. This makes it feasible to generate explanations in milliseconds for real-time fraud scoring pipelines. For deep learning models used in sequence-based fraud detection, DeepSHAP uses a linear composition rule to approximate SHAP values by backpropagating attributions through the network layers.

A precursor to SHAP that explains individual predictions by approximating the complex model locally with an interpretable surrogate model. Key differences from SHAP:

• Perturbs input data around the prediction to create a local neighborhood
• Fits a sparse linear model to the perturbed samples
• Does not guarantee the Shapley axioms of consistency and additivity
• Computationally lighter but less theoretically rigorous for feature attribution
• Often used as a faster alternative when exact Shapley values are prohibitively expensive

A global feature importance method that measures the decrease in model performance when a feature's values are randomly shuffled. Contrast with SHAP:

• Measures overall feature impact across the entire dataset, not per-prediction
• Does not reveal directionality (whether a feature pushes prediction higher or lower)
• Computationally efficient but can be misleading when features are correlated
• SHAP decomposes individual predictions; permutation importance aggregates model-level effects
• Both are model-agnostic and can be applied to any fitted estimator

A visualization technique showing the marginal effect of one or two features on the predicted outcome. Relationship to SHAP:

• PDPs average over the effects of all other features, showing a global trend
• SHAP dependence plots show the exact feature contribution per instance, colored by interaction effects
• PDPs can hide heterogeneous effects that SHAP reveals through vertical dispersion
• Combining PDP with SHAP values provides both aggregated trends and instance-level nuance
• Both assume feature independence for accurate interpretation

A polynomial-time algorithm that computes exact Shapley values for tree-based models (XGBoost, LightGBM, CatBoost, Random Forests). Advantages over Kernel SHAP:

• Exploits the internal tree structure to reduce complexity from O(2^M) to O(TLD²)
• Computes SHAP interaction values directly, quantifying pairwise feature interactions
• Returns consistent, deterministic results without sampling variance
• Integrated natively into XGBoost and LightGBM via their predict methods
• The default choice for gradient-boosted fraud detection models

An extension of SHAP that decomposes predictions into main effects and pairwise interaction effects. Key properties:

• Captures synergies where two features together influence predictions beyond their individual contributions
• The SHAP value for a feature equals its main effect plus half of all its interaction effects
• Visualized as a matrix or dependence plot with color-coded interaction strength
• Critical for fraud detection where combinations of features (e.g., high amount + unusual time) trigger alerts
• Computationally more expensive than standard SHAP values

The foundational concept from cooperative game theory that SHAP adapts. Original formulation by Lloyd Shapley (1953):

• Distributes the total payout of a coalition game fairly among players based on their marginal contributions
• Satisfies four axioms: Efficiency (sum equals total gain), Symmetry (identical contributions get equal payouts), Dummy (zero contribution gets zero), Additivity (combined games sum correctly)
• SHAP maps features to players and model prediction to the payout
• The only attribution method proven to satisfy all four axioms simultaneously
• Nobel Prize in Economics (2012) recognized Shapley's foundational work