Global Explanation Techniques

• Global explanation techniques are methods that convert the overall behavior of black-box models into human-understandable representations using rules, surrogates, and feature importance rankings.
• They employ diverse methodologies such as rule-based surrogates, game-theoretic attributions, and counterfactual recourse methods to balance fidelity and interpretability.
• These approaches are applicable across domains—from tabular data to images, text, and graphs—providing actionable insights into model decision-making.

Global explanation techniques provide interpretable, model-wide rationales for black-box machine learning models, mapping complex decision boundaries to human-understandable artifacts. Unlike local explanations—which justify individual predictions—global explanation methods aim to capture and summarize the model's behavior across the entire input space, often trading off fidelity for interpretability and scalability.

1. Fundamental Principles and Definitions

Global explanation techniques seek to translate the decision logic of a black-box model $M$ into interpretable formats that are succinct yet faithful to the model's outputs. Key objectives include:

• Fidelity: The degree to which the explanation predicts $M(x)$ for any $x$ drawn from the data distribution.
• Interpretability: The cognitive accessibility of the explanation, often quantified via the number and complexity of rules, prototypes, or features presented.
• Coverage and Robustness: The range of the input space for which the global explainer reliably reflects model behavior, and the stability of explanations under distribution shift or perturbation.

Artifact types include sets of if–then rules, symbolic logic programs, feature-importance rankings, partitioned linear models, subpopulation-specific attributions, concept-based scores, and counterfactual recourse rules.

2. Rule-Based Global Surrogates

Rule-based methods construct a set of human-interpretable logical rules or decision sets that globally approximate $M$ on tabular/classification tasks. Prominent frameworks include:

• MAGIX:
  • Combines local condition mining (via LIME) with a per-class genetic algorithm for rule induction.
  • Each rule is a conjunction of feature–value (or feature–range) conditions, encoded as a bitstring.
  • Fitness is based on mutual information between rule firing and model prediction, sign-corrected to penalize anti-correlation, and regularized by rule length.
  • Pruning eliminates redundant and non-generalizing rules, with final selection by residual overlap (Jaccard index ≤ 0.5) and fidelity constraints.
  • Achieves high imitation accuracy (typically >90% with 10 rules/class) and a favorable trade-off between interpretability and fidelity; e.g., on the Iris dataset, a one-clause rule “If petal-width ≤ 0.80 then Predict Setosa” achieves 100% precision/coverage (Puri et al., 2017).
• GA-I-TREE and similar evolutionary/exact-extraction frameworks further optimize rule compactness and stability using information-theoretic and robustness-based objectives (Verma et al., 2021).
• xDNN(ASP):
  • Decompositional extraction of symbolic logic programs representing neural network behavior under answer set semantics.
  • Each rule chains conditions across layers, capturing internal feature and hidden-node dependencies.
  • Yields both feature- and hidden-node global importance metrics; enables programmatic analysis and pruning of irrelevant nodes, supporting downstream network simplification (Trieu et al., 7 Jan 2026).

3. Global Feature-Based and Game-Theoretic Explanations

Feature-based global explanations assign importance scores to features or feature groups, often by analyzing model variance, risk, or permutation effects. Foundational approaches include:

• Functional ANOVA (fANOVA) and Cooperative Game Theory:
  • Any model $F: \mathbb{R}^d \to \mathcal{Y}$ admits a variance or effect decomposition as $F(x) = \sum_{S\subseteq D} f_S(x_S)$.
  • Variants differ by imputation: baseline (b-fANOVA), marginal (m-fANOVA), conditional (c-fANOVA), impacting sensitivity to feature interactions and data distribution.
  • Game-theoretic attributions such as the Shapley value, Sobol indices (pure, total, Shapley-variance), and interaction indices are special cases arising from this decomposition family.
  • This framework reveals that existing global feature-importance techniques can be unified as specific choices of imputation (data-centric vs model-centric) and interaction attribution (pure, partial, or full). Practitioners can thus systematically choose explanations aligned with investigative goals (variance vs performance, main effect vs interaction) (Fumagalli et al., 2024).
• Benchmarking Surveys:
  • Surrogate trees (e.g., TREPAN), rule lists (RuleFit, Bayesian Rule Lists), and additive models (EBM/GAMs) remain widely used surrogates, each with quantified fidelity–interpretability trade-offs.
  • Prototypes and concept-based summarization are prominent in image domains, but their artifact semantics and granularity differ fundamentally from rule or tree-based surrogates (Bodria et al., 2021).

4. Global Counterfactual and Recourse Summaries

Global counterfactual explanation frameworks aim to identify a small set of actions, rules, or transformations that, when applied to negatively affected populations, deliver maximal coverage of recourse (class change) with minimum average cost and sufficient interpretability.

• AReS and Improvements:
  • Constructs a compact set of recourse rules via frequent itemset mining, cross-product enumeration, and submodular maximization under interpretability constraints (number of rules, rule width, subgroup count).
  • Enforces strict global objectives on recourse coverage and cost, incorporating matroid intersection constraints and leveraging pruning for tractable ground-set evaluation.
  • Theoretical approximation guarantees via non-monotone submodular optimization; empirical speedup of up to $10^4\times$ over brute-force enumeration.
  • Demonstrated for recourse in credit and financial decision datasets (Ley et al., 2022).
• GLOBE-CE and GLANCE:
  • GLOBE-CE learns a small set of fixed translation directions (feature difference vectors) and, for each sample, computes the minimal scaling required to achieve class change. The resulting coverage–cost curve enables efficient global recourse analysis, especially in high dimensions; the methodology includes explicit theorems for categorical translation interpretability (Ley et al., 2023).
  • GLANCE (C-GLANCE and T-GLANCE algorithms) leverages clustering (in data and action space) or tree partitioning to minimize the number of global actions needed for near-complete recourse coverage at low cost—offering displayable, actionable interventions even for large populations. Outperforms previous global CE methods in both fidelity and interpretability metrics (Kavouras et al., 2024).

5. Global Explanations for Structured and Non-Tabular Domains

Specialized global explanation strategies have been devised for images, text, graphs, and time series:

• Image and Vision Models:
  • Rad4XCNN links CNN-derived features to interpretable, clinician-understood radiomic features via Spearman correlations, producing a global feature ranking that reflects the true semantic content influencing model decisions without degrading predictive accuracy. Saliency-based explanation approaches (e.g., Grad-CAM) are found insufficient for global insight due to their local nature and incoherence over datasets (Prinzi et al., 2024).
  • GEPC propagates user-annotated part labels across large datasets using deep-feature correspondence (hyperpixel flow) and aggregates part-based local minimal explanations into compact global rule lists via greedy set cover, yielding symbolic DNF (decision list) formats with high coverage and human interpretability (Rathore et al., 18 Sep 2025).
  • Concept-based methods: TCAV and its descendants yield global class–concepts scores by probing sensitivity to semantically meaningful directions in the activation space (Schrouff et al., 2021).
• Graph Neural Networks:
  • Description Logic Approach: Global explanations are constructed as class expressions (EL fragment), capturing conjunctions and existential restrictions using beam search. Scoring functions based either on model response to synthetic graphs or fidelity to the GNN on validation data optimize the expressiveness and coverage of the explanation (Köhler et al., 2024).
• Time Series:
  • L2GTX: Aggregates local explanations consisting of parameterized temporal event primitives (e.g., trend, extrema) into global, class-wise summaries via meta-clustering and budgeted instance selection, ensuring stable global faithfulness and high interpretability even under strong cluster consolidation (Mekonnen et al., 13 Mar 2026).
• Textual Models:
  • Therapy generates synthetic data through a cooperative generation protocol using MCTS-guided decoding conditioned on class-labels, then fits global feature weights via a sparse linear model, thereby identifying class–token associations with no dependence on real input data (Chaffin et al., 2023).
  • Aggregation of local anchor explanations can be globally summarized via accelerated algorithms—using anytime selection, upper bounding, and proprietary noise models to robustly identify globally impactful words/features even in the presence of token frequency bias (Mor et al., 2023).
  • Relevance thesaurus: In neural ranking (IR) tasks, a massive term–term “relevance thesaurus” is distilled by training partial-input neural models. The thesaurus recovers much of the neural model's behavior in lexical models and exposes global biases not discoverable from local attributions (Kim et al., 2024).

6. Subpopulation and Partitioned-Local Global Explanations

Clustering-based and hybrid strategies go beyond simple global feature importances to capture model heterogeneity across subspaces:

• Global Attribution Mapping (GAM):
  • Aggregates normalized local attributions into $K$ clusters (medoids) using rank distance (weighted Kendall/Spearman), reporting both global attributions (mean/medoid) and the subpopulation fractions each best describes.
  • Tunable $K$ controls granularity, yielding transparent mappings between sample groups and global explanations; this addresses limitations of single global importance vectors and is empirically validated against known ground truth and human acceptance (Ibrahim et al., 2019).
• GLEAMS:
  • Learns a global partition of the input space (hyper-rectangle) via recursive, score-based splitting, fitting an OLS linear model in each cell so that both global surrogacy and on-demand, locally faithful linear attributions are available.
  • Guarantees per-cell fidelity and controlled complexity; outperforms both global surrogates and aggregated local explainers in fidelity–interpretability trade-offs (Visani et al., 2024).

7. Strengths, Limitations, and Theoretical Guarantees

Strengths	Limitations	Representative Methods
High global fidelity with interpretable rules	May require significant computation (GA, clustering)	MAGIX, xDNN(ASP), AReS, GAM
Tunable complexity and trade-off mechanisms	Vulnerable to spurious or unstable local feature picks	fANOVA–Shapley, GLEAMS, L2GTX
Domain-specific semantic mappings (e.g. parts, concepts, radiomics)	Need for pre-labeled concepts/parts in some frameworks	GEPC, TCAV, Rad4XCNN
Robustness to distribution shift (with OOD augmentation)	Varying sensitivity to data imputation choices	MAGIX, fANOVA
Human-auditability and interactive support	Categorical/structured features still challenging	GLANCE, GLOBE-CE, IR thesaurus

Theoretical guarantees encompass fidelity bounds (often measured as proportion of agreement between surrogate and black-box), approximation ratios for submodular optimization under interpretability constraints (AReS, up to $1/(k+1)$ for $M(x)$0 matroid constraints), and stratified program correctness for logic program (ASP)–based extraction. Trade-offs between interpretability, expressiveness, and computational feasibility remain central in the comparative evaluations.

8. Conclusion and Outlook

Global explanation techniques represent a rich and active research area, distinguished by the diversity of their artifacts (rules, surrogates, partitions, concepts, actions), performance metrics (fidelity, coverage, cost), and domain targeting (tabular, image, text, graph, time series). Technique selection should consider the desired semantic level (feature, concept, action), interpretability constraints (rule set size, prototype coverage), and domain-specific needs (e.g., explainability–accuracy trade-off in clinical imaging, subgroup recourse in financial fairness).

Ongoing research aims to close the gaps in cross-domain generality, enhance robustness under distribution shift, enable scalable partitioning for structured data, and support hybrid explanations combining local, global, and concept-driven rationales (Puri et al., 2017, Verma et al., 2021, Trieu et al., 7 Jan 2026, Fumagalli et al., 2024, Prinzi et al., 2024, Rathore et al., 18 Sep 2025, Ley et al., 2022, Mor et al., 2023, Mekonnen et al., 13 Mar 2026, Kim et al., 2024, Ley et al., 2023, Visani et al., 2024, Ibrahim et al., 2019, Köhler et al., 2024, Schrouff et al., 2021, Chaffin et al., 2023, Kavouras et al., 2024, Bodria et al., 2021).