SHAP 值：全局与局部特征重要性

Explainability is the ability to understand why a model made a specific prediction. In high-stakes domains like lending, healthcare, and hiring, regulators and users demand explanations — not just accurate predictions. SHAP (SHapley Additive exPlanations) provides a mathematically principled framework rooted in cooperative game theory to deliver these explanations for any model.

SHAP values borrow from Shapley values in cooperative game theory, where players (features) collaborate to produce an outcome (prediction). Each feature receives a fair share of credit by averaging its marginal contribution across all possible feature orderings. This makes SHAP the only additive attribution method satisfying the axioms of efficiency, symmetry, dummy, and additivity.

The shap library supports tree models, neural networks, and any black-box model. Install it with pip install shap and import it alongside your trained model. SHAP's explainers are model-type-aware: TreeExplainer for gradient boosting and random forests gives exact values in O(T·D²) time, far faster than the naive exponential-time Shapley computation.

Call explainer.shap_values(X_test) to produce a matrix where each row is a sample and each column is a feature. The SHAP value for feature j in sample i represents the contribution of feature j to pushing the prediction away from the expected base value. Positive values push toward the positive class; negative values push toward the negative class.

Global feature importance summarises which features matter most across all predictions. The SHAP summary_plot in bar mode shows the mean absolute SHAP value per feature, ranking them from most to least important. This replaces the naive built-in feature importance that only counts split counts, which is biased toward high-cardinality features.

The beeswarm plot (default summary_plot) is richer than a bar chart: each dot represents one sample, coloured by feature value (red = high, blue = low). The x-axis shows the SHAP value, so you can see not only which features matter but also in which direction a high or low feature value pushes predictions. This reveals nonlinear and interaction effects at a glance.

A force plot explains a single prediction. It shows the base value on the left and the final prediction on the right, with features as arrows that push the output higher (red) or lower (blue). The width of each arrow is proportional to the feature's SHAP value. This is the explanation you would show a loan officer asking 'why was this application denied?'

The waterfall plot is a cleaner alternative to the force plot for a single sample. It stacks SHAP contributions vertically from the base value, showing each feature's contribution as a bar segment. Positive contributions are red and push toward the top; negative contributions are blue and pull down. The final stack total equals the model's raw output for that sample.

A SHAP dependence plot shows how a single feature's SHAP value changes as its raw value changes, coloured by a second feature to reveal interactions. For example, plotting 'worst radius' coloured by 'mean texture' reveals whether the effect of radius depends on texture. This goes beyond ordinary partial-dependence plots by accounting for all feature interactions naturally.

When the model is a black box (SVM, neural network, any sklearn estimator), use shap.KernelExplainer, which approximates Shapley values by sampling coalitions and fitting a weighted linear model locally. It is model-agnostic but slower than TreeExplainer. Provide a background dataset summary (e.g., K-Means centroids) to speed up computation on large datasets.

A key property of SHAP is efficiency: the sum of all SHAP values for a sample plus the base value must equal the model's raw output. Verifying this sanity check confirms the explainer is working correctly. Any discrepancy indicates a mismatch between the explainer type and the model, or incorrect background data.

In this lesson you learned: SHAP values quantify each feature's contribution to a prediction using Shapley values from game theory, global summaries (bar and beeswarm plots) reveal overall feature importance and direction, and local explanations (force and waterfall plots) justify individual predictions. Next up we explore LIME as an alternative model-agnostic explanation approach.