xai-auditor

Quantitative Explainability Auditing and Model Risk Management for Machine Learning.

xai-auditor is an independent, production-grade Python library designed to audit the reliability, stability, fidelity, consistency, manipulation resistance, distributional drift, and operational coverage of post-hoc Explainable AI (XAI) pipelines such as SHAP and LIME.

Overview

In regulated and high-stakes machine learning applications, post-hoc explanations must be audited with the same mathematical rigor applied to model predictive accuracy. xai-auditor evaluates whether surrogate explainers produce faithful, robust, and dependable explanations across six core dimensions:

1. Local Lipschitz Stability: Attribution invariance and rank preservation under small feature perturbations.
2. Attribution Fidelity: Empirical causal validity measured via ROAR (Remove and Retain) and KAR (Keep and Retain) probability curves.
3. Cross-Method Explainer Consistency: Consensus between SHAP and LIME evaluated via Top-K Jaccard overlap, Spearman rank correlation, Sign Agreement, Cosine Similarity, and Rank-Biased Overlap (RBO).
4. Adversarial Manipulation Resistance: Robustness against explainer hyperparameter tuning (kernel widths, sampling budgets, background reference sets) measured by the Explanation Manipulation Index (EMI).
5. Distributional Explanation Drift: Temporal and cohort attribution shift evaluated via Population Stability Index (PSI) and Wasserstein Distance.
6. Operational Coverage & Latency: Service level agreement (SLA) verification evaluating explanation success rates and p50/p95/p99 execution latency.

Installation

pip install xai-auditor

Optional dependencies for deep integration with NumPy, Pandas, Scikit-Learn, SHAP, and LIME:

pip install xai-auditor[all]

Quickstart

from xai_auditor import Auditor
from xai_auditor.models.builtins import BuiltinLogisticRegression

# 1. Prepare data and model
# X can be a list of lists, a numpy array, or a pandas DataFrame
X = [
    [0.2, 1.1, 0.4, 0.9],
    [0.8, 0.3, 0.1, 0.2],
    [0.5, 0.9, 0.7, 0.6],
    [0.1, 0.2, 0.8, 0.4],
    [0.9, 1.4, 0.3, 0.8],
    [0.3, 0.5, 0.2, 0.1],
]
y = [1, 0, 1, 0, 1, 0]
feature_names = ["credit_utilization", "annual_income", "debt_ratio", "payment_history"]

model = BuiltinLogisticRegression(weights=[0.8, -0.6, 1.2, -0.9], bias=0.1)

# 2. Run the audit
auditor = Auditor(
    model=model,
    X=X,
    y=y,
    feature_names=feature_names,
)
result = auditor.run()

# 3. Inspect scores, risk tier, and metrics
print(f"Explainability Trust Score: {result.trust_score}/100")
print(f"Risk Classification Tier:   {result.risk_tier}")
print(f"Regulatory Approval Status: {result.approval_status}")

# 4. Access individual dimension scores
for dim, dim_result in result.dimension_results.items():
    print(f" - {dim_result.name}: {dim_result.score}/100 (Weight: {int(dim_result.weight * 100)}%)")

# 5. Export results to JSON or HTML
json_str = result.to_json(indent=2)
result.save_html_report("audit_report.html")

Auditing Individual Dimensions

Users can run individual audit dimensions independently without running the full suite:

from xai_auditor import (
    StabilityAuditor,
    FidelityAuditor,
    ConsistencyAuditor,
    ManipulationAuditor,
    DriftAuditor,
    CoverageAuditor,
)

# Example: Run only cross-method consistency (SHAP vs LIME)
consistency_auditor = ConsistencyAuditor(top_k=3, rbo_p=0.9)
res = consistency_auditor.audit(model=model, X=X, feature_names=feature_names)

print(f"Consistency Score: {res.score}/100")
print(f"Top-K Jaccard:    {res.metrics['top_k_jaccard']}")
print(f"Sign Agreement:   {res.metrics['sign_agreement']}")
print(f"Rank-Biased RBO:  {res.metrics['rank_biased_overlap']}")

Supported Models and Explainers

Models

• Scikit-Learn classifiers implementing predict_proba
• XGBoost, LightGBM, CatBoost binary classifiers
• PyTorch / TensorFlow binary classification modules
• Built-in standalone reference models (XGBoost emulation, Random Forest, Logistic Regression, Multilayer Perceptron)
• Arbitrary Python prediction callables: f(X) -> probabilities

Explainers

• SHAP: Permutation Shapley engine with efficiency constraint enforcement
• LIME: Kernel-weighted local linear ridge surrogate with exponential distance weighting

Scoring and Risk Classification

The composite Explainability Trust Score (0-100) aggregates all six dimensions using established regulatory weights:

Risk Tiers

• Low Risk (85 - 100): Approved for unconditional production deployment.
• Medium Risk (75 - 84): Approved with standard governance telemetry.
• High Risk (60 - 74): Conditional approval with mandatory secondary review.
• Critical Risk (0 - 59): Deployment rejected; hard re-audit mandated.

Command-Line Interface (CLI)

xai-auditor provides a command-line tool xai-audit for batch auditing and continuous integration:

xai-audit --data data.csv --target is_default --model logistic_regression --output report.json --html report.html

License

Apache License 2.0.