Shapley Attribution
Computes Shapley values to quantify feature contributions to machine learning model predictions. Supports exact and approximate attribution methods for black-box models. Data scientists and ML engineers use it to generate interpretable explanations in deployment pipelines.
Overview
The Shapley Attribution MCP server implements Shapley value calculations for feature attribution in machine learning models. Derived from cooperative game theory, Shapley values provide a fair allocation of a model's prediction to individual input features, satisfying properties like efficiency and symmetry. This server exposes these computations via MCP tools, enabling integration into analysis workflows without custom implementation.
Key Capabilities
- shapley_attribution: Calculates Shapley values for a given model, dataset, and target prediction. Supports permutation sampling for approximations on large feature sets and exact computation for smaller ones. Inputs include model object, feature matrix, and baseline values; outputs a vector of attributions per feature.
The server handles model-agnostic attribution, working with tree-based, neural, or tabular models.
Use Cases
-
Model Debugging: Run shapley_attribution on a fraud detection model to identify if transaction amount or location drives high-risk scores, aiding retraining decisions.
-
Regulatory Compliance: In finance, compute attributions for credit scoring models to document feature impacts per loan application, meeting explainability mandates.
-
Healthcare Diagnostics: Apply to patient risk models, attributing predictions to vitals, labs, and demographics for clinician review.
-
A/B Testing Analysis: Compare feature contributions across model versions to quantify performance deltas.
Who This Is For
ML engineers deploying interpretable models, data scientists analyzing prediction drivers, and researchers in explainable AI (XAI). Suited for teams using frameworks like scikit-learn, XGBoost, or PyTorch needing production-grade attribution.