Intrinsic Interpretability

The model's prediction is a weighted sum of input features, where each coefficient directly quantifies the feature's contribution. In fraud detection, a logistic regression coefficient of +0.8 for transaction_amount_z_score means a one-unit increase in the z-score multiplies the odds of a fraud flag by e^0.8.

• Direct magnitude interpretation: Larger absolute coefficients indicate stronger influence
• Sign indicates direction: Positive coefficients increase fraud probability; negative coefficients decrease it
• Statistical significance: P-values and confidence intervals provide uncertainty quantification
• No interaction effects: Assumes each feature acts independently, which is both a strength for interpretability and a limitation for capturing complex fraud patterns

A single decision tree makes predictions by traversing a series of binary splits from root to leaf. Each prediction can be explained by listing the exact sequence of feature thresholds that were satisfied. For a transaction flagged as fraudulent, the explanation might read: amount > $10,000 → merchant_category = 'jewelry' → time_since_last_transaction < 5 minutes → Fraud Score: 0.92.

• Rule extraction: Each path from root to leaf is a human-readable conjunctive rule
• Feature importance via split frequency: Features used closer to the root have greater global impact
• No feature scaling required: Trees are invariant to monotonic transformations
• Limitation: Single trees are high-variance and prone to overfitting without ensemble methods

EBMs are generalized additive models with pairwise interactions, where each feature contributes through a learned shape function f(x_i). The final prediction is the sum: f_1(x_1) + f_2(x_2) + ... + f_{ij}(x_i, x_j). Each shape function can be plotted and inspected independently, showing exactly how a feature like account_age_days influences fraud risk across its range.

• Modular interpretability: Each feature's contribution is isolated and visualizable
• Pairwise interaction detection: Automatically identifies and models feature interactions while keeping them interpretable
• Monotonicity constraints: Can enforce that risk must increase with certain features (e.g., transaction velocity)
• Performance parity: Often matches gradient boosted trees while remaining fully transparent

Rule-based models consist of explicit if-then statements that map directly to domain knowledge. In anti-money laundering, a rule might be: IF cash_deposit > $10,000 AND account_opened < 30 days THEN flag_for_SAR_review. Every decision is traceable to a specific, named rule with a clear business justification.

• Zero ambiguity: Every decision is explained by the exact rule that fired
• Regulatory alignment: Rules can be mapped directly to compliance requirements (BSA, AML directives)
• Auditability: Rule changes are version-controlled and require explicit approval
• Limitation: Cannot detect novel fraud patterns not anticipated by rule authors; requires continuous manual updates

NAMs learn a neural network for each feature independently, producing shape functions f(x_i) that are then summed for the final prediction. Unlike standard deep networks, each feature's learned function can be plotted as a curve showing exactly how the model's output changes with that feature. For transaction_frequency_7d, the plot might show a sharp risk increase after 15 transactions.

• Deep learning expressiveness with linear interpretability: Each shape function can capture complex non-linearities
• Feature-level regularization: Can apply different smoothness constraints per feature
• Statistical testing: Shape functions can be tested for significance like linear coefficients
• Training efficiency: Each feature network trains independently before final joint calibration

Intrinsically interpretable models can enforce monotonic relationships between inputs and outputs, ensuring that risk predictions never decrease when risk factors increase. A fraud model with monotonicity constraints guarantees that a higher failed_login_attempts count never results in a lower fraud score, satisfying regulatory requirements for logically consistent decision-making.

• Regulatory defensibility: Prevents counterintuitive decisions that undermine model credibility
• Implementation via constrained optimization: Enforced during training through architectural constraints or penalty terms
• Applicable models: EBMs, NAMs, and constrained logistic regression support monotonicity natively
• Audit simplification: Eliminates entire categories of model risk findings related to directional inconsistency

The application of interpretation methods to a trained, complex model after it has been built. Unlike intrinsic interpretability, which relies on simple structures like linear regression, post-hoc techniques such as SHAP and LIME are used to decode opaque neural networks. This is critical for fraud detection systems where high accuracy from black-box models must be reconciled with regulatory demands for adverse action reason codes.

An interpretable model, such as a decision tree or linear regression, trained to approximate the predictions of a complex black-box model. This provides global insight into the black-box's behavior. For instance, a surrogate model can reveal that a deep learning fraud detector is primarily using transaction velocity and geolocation mismatch to flag anomalies, mimicking the transparency of an intrinsically interpretable system.

A glass-box model that combines the accuracy of gradient boosting with the transparency of generalized additive models. EBM learns interpretable shape functions for each input feature, allowing for exact, modular explanations. In fraud detection, an EBM can show precisely how a specific transaction amount or time since last login contributes to a risk score, providing the auditability of intrinsic models without sacrificing performance.

A set of concise, human-readable statements that provide the primary drivers behind a specific decision. For intrinsically interpretable models, reason codes are derived directly from model coefficients or tree splits. Regulations like the Fair Credit Reporting Act (FCRA) mandate adverse action reason codes when a transaction is blocked, making the direct extraction of these codes a key advantage of using inherently simple models in financial applications.

A comprehensive, chronological record of the data, model parameters, and logic used for a specific transaction. Intrinsically interpretable models simplify the creation of audit trails because their decision logic is explicit. An auditor can trace a fraud score directly back to a rule like IF amount > $10,000 AND account_age < 30 days THEN high_risk, ensuring full traceability and accountability for regulatory review.

A process of transferring knowledge from a large, complex teacher model to a smaller, more efficient student model. Often, the student is an intrinsically interpretable structure like a decision tree or a Neural Additive Model (NAM). This allows a fraud detection system to leverage the high accuracy of a deep learning teacher while deploying a transparent, low-latency student model that can be directly audited.