shap – skill Bảo mật cho agent AI

SHAP

Use SHAP to describe how a fitted predictive model maps inputs to outputs. Work from the modern shap.Explanation API, make the explained output and background distribution explicit, and validate every explanation before interpreting it.

This skill is aligned with SHAP 0.52.0 (released 2026-05-28). That release requires Python 3.12 or newer.

The maintained examples were checked on Python 3.12.10 with SHAP 0.52.0, NumPy 2.5.3, pandas 3.0.6, scikit-learn 1.9.1, and matplotlib 3.11.2. Native tests cover small tree, exact, permutation, partition, linear, additive, kernel, text, and constant-image games; additional numeric XGBoost 3.4.1 smoke checks cover raw, probability, loss, and interaction outputs. Reference snippets that require a project model, framework, or data are adaptation templates; optional pretrained/deep, GPU, and distributed integrations remain illustrative and require their own runtime validation.

Operating Rules

1. Explain a fixed, evaluated model; do not use SHAP as a substitute for predictive validation.

2. Use held-out or clearly labeled analysis rows for explanations. Choose background rows only from an appropriate training or reference population.

3. State the explained output: regression value, raw margin, probability, log loss, logit, or another model method.

4. Prefer shap.Explanation objects and explainer(X). Some specialized options still require .shap_values(X), including deep ranked outputs, gradient sampling budgets, and Kernel SHAP nsamples. Preserve their output indexes and baselines explicitly.

5. For multi-output models, select one output before using tabular plots: explanation[..., output_index].

6. Check base_values + values.sum(...) against the exact model output being explained.

7. Treat SHAP as a description of model behavior under a masking/background choice. It does not establish causality, fairness, recourse, or scientific mechanism.

8. Never silence an additivity failure until input shape, preprocessing, model version, output space, and row ordering have been checked.

9. Do not load untrusted pickle, joblib, model, or explainer artifacts; those formats can execute code during deserialization.

Install

Create an isolated environment and pin the documented release:

uv venv --python 3.12
source .venv/bin/activate
uv pip install "shap[plots]==0.52.0"

shap[plots] installs the plotting dependencies. Add the fitted model's package at a version compatible with the project. For older Python compatibility, read [references/migration.md](references/migration.md) instead of silently installing a different SHAP release.

Confirm the environment before debugging an API mismatch:

import platform
import shap

print("Python:", platform.python_version())
print("SHAP:", shap.__version__)

Standard Workflow

1. Define the explanation target

Record:

• model and preprocessing version;
• exact callable or model method being explained;
• output name/index and units;
• evaluation rows;
• background/reference population;
• masker and explainer algorithm;
• SHAP and model-library versions.

For classifiers, decide whether the task needs raw margins or probabilities. Defaults differ by model family; never infer units from the plot color or sign.

2. Select an explainer and masker

Start with shap.Explainer(model, masker) when automatic dispatch is sufficient. Instantiate a specialized explainer when its assumptions or output controls matter.

| Situation | Preferred choice | Important constraint |

|---|---|---|

| Supported tree ensemble | TreeExplainer | model_output="probability" and "log_loss" require interventional masking and background data |

| Linear model | LinearExplainer | The masker determines interventional versus correlation-aware behavior |

| Small feature space | ExactExplainer | Cost grows quickly with unconstrained feature count |

| General tabular callable | PermutationExplainer | Budget at least one full forward/reverse permutation |

| Hierarchical feature groups, text, or image | PartitionExplainer | The partition tree changes the cooperative game |

| Differentiable neural network | DeepExplainer or GradientExplainer | Framework support, output shape, and background choice require testing |

| Legacy Kernel SHAP workflow | KernelExplainer | Usually much slower than model-specific methods |

Use the detailed decision guide in [references/explainers.md](references/explainers.md). Use [references/data-maskers.md](references/data-maskers.md) when features are correlated, structured, sparse, or semantically grouped.

3. Compute a modern Explanation

This complete binary-classification example uses an explicit background and selects output index 1. In the breast-cancer dataset, class 1 means benign, so positive SHAP values below increase predicted benign probability, not cancer risk. For another dataset, resolve the requested label through model.classes_ and record its meaning before selecting an output index; column 1 is not universally the clinically positive event.

import numpy as np
import shap
from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

X, y = load_breast_cancer(as_frame=True, return_X_y=True)
X = X.astype(np.float32)  # Match sklearn forest prediction inputs.
X_train, X_test, y_train, y_test = train_test_split(
    X,
    y,
    test_size=0.25,
    stratify=y,
    random_state=7,
)

model = RandomForestClassifier(
    n_estimators=200,
    min_samples_leaf=3,
    random_state=7,
    n_jobs=-1,
).fit(X_train, y_train)

X_test = X_test.iloc[:100]  # Predeclared held-out explanation subset.
background = shap.sample(X_train, 100, random_state=7)
explainer = shap.Explainer(model, background, algorithm="tree")
all_outputs = explainer(X_test)

# sklearn tree classifiers expose one output per class.
positive = all_outputs[..., 1]
assert positive.values.shape == X_test.shape

reconstructed = np.asarray(positive.base_values) + positive.values.sum(axis=1)
expected = model.predict_proba(X_test)[:, 1]
np.testing.assert_allclose(reconstructed, expected, rtol=1e-5, atol=1e-6)

shap.plots.beeswarm(positive, max_display=15)
shap.plots.waterfall(positive[0], max_display=15)

Output shape is model-dependent:

• one tabular output: (samples, features);
• multiple tabular outputs: (samples, features, outputs);
• multiple model inputs: often a list of arrays or explanations;
• image/text explanations: feature axes follow the input representation, with output selection on the final axis when present.

Do not use the pre-0.45 pattern values[class_index] for a modern multi-output array. Use values[..., class_index] or slice the Explanation itself.

4. Control tree output semantics when needed

For a supported tree classifier, probability-space explanations must be explicit:

background = shap.sample(X_train, 200, random_state=7)

explainer = shap.TreeExplainer(
    model,
    data=shap.maskers.Independent(background, max_samples=len(background)),
    feature_perturbation="interventional",
    model_output="probability",
)
probability_exp = explainer(X_test)

A bare background frame is capped at 100 rows by the default masker. The explicit masker above retains all 200 sampled rows; inspect len(explainer.data) when reporting or comparing background sizes.

In SHAP 0.52:

• feature_perturbation="auto" uses interventional semantics when background data is supplied and tree-path-dependent semantics otherwise;
• probability and log-loss output modes are supported only with interventional semantics;
• pass approximate=True to explainer(X, approximate=True) if deliberately using the lower-fidelity tree approximation; do not pass it to the constructor.

5. Use a model-agnostic callable deliberately

Pass the exact callable whose outputs will be interpreted:

masker = shap.maskers.Independent(background, max_samples=100)
explainer = shap.Explainer(
    model.predict_proba,
    masker,
    algorithm="permutation",
    output_names=[str(label) for label in model.classes_],
    seed=7,
)

budget = 2 * X_test.shape[1] + 1
all_outputs = explainer(X_test.iloc[:20], max_evals=budget)
# 0.52 selector dispatch may drop output_names for permutation.
all_outputs.output_names = [str(label) for label in model.classes_]
positive = all_outputs[..., 1]

Increase max_evals to average over more permutations when estimates are unstable. Keep the seed, background sample, and evaluation budget in the report.

6. Visualize the question, not merely the available plot

| Question | Plot |

|---|---|

| Which features have the largest average attribution magnitude? | shap.plots.bar(exp) |

| How do direction, magnitude, and observed values vary globally? | shap.plots.beeswarm(exp) |

| Why did one prediction differ from its baseline? | `shap.plots.waterfall(exp[i])

… (nội dung đầy đủ có trong gói tải về)