Explainable Stacking Ensemble
Explainable Stacking Ensemble (Interpretable Meta-Learning) · Also known as: XAI-Stacking, interpretable stacking, transparent stacking ensemble, explainable stacked generalisation
Explainable Stacking Ensemble combines the predictive power of stacked generalisation — training a meta-learner on the outputs of multiple diverse base models — with interpretability tools such as SHAP or LIME that reveal how each base model and each input feature contributed to the final prediction. It bridges the accuracy–transparency trade-off that makes pure stacking opaque in high-stakes settings.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Explainable Stacking Ensemble when you need top predictive accuracy on tabular data and must also justify individual predictions to stakeholders, regulators, or scientific reviewers — for example in clinical risk scoring, credit assessment, or policy-relevant prediction. It suits datasets with at least a few hundred observations and several candidate models of different types. Do not use it when a single interpretable model (logistic regression, decision tree) is sufficient for the task, when computational budget is tight, or when the audience has no tolerance for model ensembles. Avoid it if you need a single closed-form equation rather than a pipeline.
Strengths & limitations
- Combines the best predictions of diverse base models while retaining interpretability via SHAP or LIME.
- Flexible: any set of base learners can be assembled, allowing domain knowledge to guide model selection.
- Meta-learner contribution scores provide a principled answer to which sub-model matters most.
- Out-of-fold training protocol prevents data leakage and gives unbiased meta-features.
- Widely applicable to classification and regression on mixed tabular data without distributional assumptions.
- Substantially more complex to implement, validate, and communicate than a single model.
- Explanation quality depends on the fidelity of the post-hoc tool: SHAP is not exact for all meta-learners.
- Training cost scales with the number of base learners and cross-validation folds.
- Propagating explanations from the meta-learner back to raw features is an approximation, not a true end-to-end derivation.
- Reproducibility requires careful versioning of all base learners, their OOF predictions, and the meta-learner.
Frequently asked
How is explainable stacking different from a plain stacking ensemble?
A plain stacking ensemble optimises for accuracy. Explainable stacking adds an interpretability layer — usually SHAP values on the meta-learner and base learners — so that each prediction can be decomposed into contributions from individual models and features. The predictive pipeline is the same; the transparency layer is added on top.
Which meta-learner is best for interpretability?
Logistic regression (for classification) or linear regression (for regression) are the most transparent meta-learners because their coefficients directly reflect each base model's weight. Shallow decision trees are also common. Complex meta-learners like gradient boosting can improve accuracy but require additional SHAP analysis to interpret.
Does adding SHAP explanations change the predictions?
No. SHAP is a post-hoc attribution method — it analyses the already-trained model without modifying it. The predictions are identical whether or not you compute SHAP values afterwards.
How many base learners are typically used?
Three to seven diverse base learners is a common range. More learners can improve accuracy marginally but increase training cost and communication complexity. The key is diversity of inductive bias, not quantity.
Is explainable stacking suitable for small samples?
Below roughly 200 observations the cross-validation folds used to generate OOF meta-features become small, base-learner variance increases, and SHAP estimates become unstable. For small samples a single interpretable model is usually preferable.
Sources
- Wolpert, D. H. (1992). Stacked generalization. Neural Networks, 5(2), 241–259. DOI: 10.1016/S0893-6080(05)80023-1 ↗
- Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30. link ↗
How to cite this page
ScholarGate. (2026, June 3). Explainable Stacking Ensemble (Interpretable Meta-Learning). ScholarGate. https://scholargate.app/en/machine-learning/explainable-stacking-ensemble
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bagging EnsembleEnsemble Learning↔ compare
- Gradient BoostingMachine learning↔ compare
- Random ForestMachine learning↔ compare
- XGBoostMachine learning↔ compare