Explainable AI (XAI) Methods

• Explainable AI (XAI) methods are techniques that clarify complex ML models through inherent and post-hoc approaches, ensuring transparency and trust.
• These methods are categorized by characteristics such as model-agnostic vs. model-specific, local vs. global scope, and explanation format.
• Effective XAI evaluation benchmarks focus on fidelity, robustness, and complexity, guiding method selection across various application domains.

Explainable Artificial Intelligence (XAI) methods provide a framework for interpreting, understanding, and validating the outputs of complex machine learning models. XAI is increasingly essential in critical applications—medicine, finance, autonomous systems—where the opacity of predictive models poses risks for safety, fairness, and regulatory compliance. The proliferation of XAI techniques reflects a taxonomy spanning inherently interpretable "glass-box" models, post-hoc model-agnostic explainers, model-specific saliency methods, and emerging multimodal, semantic, and human-centered approaches.

1. Taxonomy of XAI Methods

XAI methods are naturally organized along four principal axes: (1) model-agnostic vs. model-specific, (2) post-hoc vs. ante-hoc (intrinsic), (3) local vs. global scope, and (4) explanation format (feature attribution, rule, visual, case-based, etc.) (Islam et al., 2021, Mumuni et al., 17 Jan 2025, Karim et al., 2022, Rojat et al., 2021, Gohel et al., 2021).

Model-agnostic post-hoc methods generate explanations by probing model inputs and outputs, producing local (individual sample) or global (model-wide) attributions without needing access to the model's internal structure. Canonical examples include LIME, SHAP, PDP, ICE, ALE, and permutation feature importance.

Model-specific methods exploit architectures such as neural networks, using gradients, feature activations, or attention weights to analyze internal decision pathways. This category includes saliency maps, Grad-CAM, Integrated Gradients, Layer-wise Relevance Propagation, and attention visualization.

Intrinsic (ante-hoc) models embed explainability into the model structure itself, yielding decisions that are inherently interpretable. Examples include linear and logistic regression, decision trees, rule-based learners, generalized additive models (GAMs), and more recently, Neural Additive Models (NAMs) and Explainable Boosting Machines (EBMs).

Local vs. global explanations: Local methods address "Why did the model predict y for this x?" (e.g., LIME, SHAP instance explanations, Anchors, counterfactuals). Global methods answer "What general patterns does the model capture?" (e.g., aggregate SHAP, PDPs, rule extraction, feature importance rankings).

2. Mathematical Formalisms and Algorithmic Workflows

Many XAI methods leverage rigorous mathematical foundations, especially for local feature attributions:

• Shapley Additive Explanations (SHAP):

For a model $f$, features $F$ of size $M$, and instance $x$:

$$\phi_i(f, x) = \sum_{S \subseteq F \setminus \{ i \} } \frac{|S|! (M-|S|-1)!}{M!} \left[ f_{S \cup \{i\}}(x_{S \cup \{i\}}) - f_S(x_S) \right]$$

The explanation is additive: $f(x) = \phi_0 + \sum_{i=1}^M \phi_i$, where $\phi_0$ is the expected output with no features. SHAP satisfies efficiency (additivity), symmetry, dummy, and linearity properties (Salih et al., 2023, Duell, 2021).

• Local Interpretable Model-agnostic Explanations (LIME):

LIME fits an interpretable surrogate $g$ (e.g. sparse linear model) to perturbations $z$ in the neighborhood of $x$, weighted by proximity $\pi_x(z)$:

$$\underset{g \in G}{\arg \min} \ \mathcal{L}(f, g, \pi_x) + \Omega(g)$$

where the loss $\mathcal{L}$ penalizes lack of fidelity to $f$, and $\Omega(g)$ controls surrogate complexity (Salih et al., 2023, Islam et al., 2021).

• Anchors:

Find minimal predicate sets $A(x)$ such that the model output $f(z)$ is invariant with high probability under perturbations $z$ constrained by $A$, i.e.:

$$\mathbb{E}_{z \sim D_x(z|A)} [ 1_{f(z)=f(x)} ] \geq \tau$$

• Integrated Gradients (IG):

For differentiable $f$ and baseline input $x'$, IG for feature $i$ is:

$$IG_i(x) = (x_i - x'_i) \int_{\alpha = 0}^1 \frac{\partial f(x' + \alpha (x - x'))}{\partial x_i} d\alpha$$

IG satisfies sensitivity and implementation invariance (Jogani et al., 2022, Bommer et al., 2023, Mumuni et al., 17 Jan 2025).

• Explainable Boosting Machines (EBM):

EBMs generalize GAMS with pairwise interactions:

$$f(x) = \beta_0 + \sum_j f_j(x_j) + \sum_{j<k} f_{jk}(x_j, x_k)$$

Each $f_j$ is a shape function learned nonparametrically, typically via small trees or splines (Duell, 2021, Mumuni et al., 17 Jan 2025).

• Counterfactual Explanations:

Minimize input perturbation subject to forced output:

$$x' = \arg\min_{x'}\ \|x'-x\| \quad \text{subject to}\ f(x') = y'$$

3. Comparative Benchmarks and Evaluation Metrics

Robust XAI benchmarking requires multidimensional evaluation. The key axes are fidelity, stability/robustness, complexity/conciseness, and randomization/sanity (Stassin et al., 2023, Bommer et al., 2023).

Metric	Property Evaluated	Typical Quantification
Fidelity	Faithfulness to $f$	Correlation between attribution and prediction drop upon feature perturbation; local/global fidelity
Stability/Robustness	Consistency	Sensitivity (mean/maximum change in attributions under small input noise); local Lipschitz estimate
Complexity/Conciseness	Human comprehensibility	Sparseness (Gini), entropy (Shannon), effective #nonzero attribution components
Randomization/Sanity	Non-triviality	Parameter randomization test, random logit test (class swapping)

Baseline choice (e.g., pixel replacement value, feature masking strategy) strongly impacts faithfulness metrics; using multiple baselines is recommended (Stassin et al., 2023). Dummy explanations (random, Sobel, Gaussian) must be included to confirm that metrics demote trivial or non-informative maps.

Instance-level agreement among XAI methods can be limited. For example, SHAP and LIME agreed on the top-1 feature in ~85% of cases in a mortality task, but top-3 agreement dropped to ~50%, highlighting reliability concerns (Duell, 2021). Effectiveness for human judgment varies by context: SHAP exhibits highest statistical fidelity, while Anchors maximize human understandability in tabular domains, and Grad-CAM/LRP dominate trust in image-based tasks (Nayebi et al., 2022, Labarta et al., 14 Oct 2024).

4. Application Contexts and Method Selection

No single XAI method is universally optimal. Selection must be aligned to model class, data modality, explanation scope, and user requirements:

• Tabular clinical data: SHAP (fidelity, stability), Anchors (transparency, local actionability), LIME (local quick looks; less stable), global surrogates for audit (Duell, 2021, Nayebi et al., 2022, Islam et al., 2021).
• Medical imaging: Grad-CAM (fast, coarse heatmaps), IG (pixel-level relevance, axiomatic), LIME (superpixel masks, stakeholder-friendly), LRP (high-resolution backprop), combined with expert validation (Jogani et al., 2022, Sadeghi et al., 2023).
• Time series and sequential data: LIME/SHAP adapted to windowed features; attention visualization and shapelet extraction for global trends (Rojat et al., 2021).
• Bioinformatics: SHAP for identifying discriminative genes, Grad-CAM++ and LRP for bioimage saliency, combined with knowledge graph lookups for semantic enrichment (Karim et al., 2022).
• High-dimensional EHR: Combine glass-box (GAM, EBM) models for global interpretability with SHAP/LIME/Anchors layered atop performant black-boxes to support local and global reasoning (Duell, 2021).

Practitioners are advised to calibrate explanation complexity to end-user expertise. For non-technical users, numerical confidence cues and category-level rationales may be more effective than composite heatmaps (Labarta et al., 14 Oct 2024).

5. Ethical, Regulatory, and Human-Centered Considerations

XAI's principal goal is to establish trust and actionable insight, particularly in safety-critical and regulated environments. Key desiderata, as framed by Tonekaboni et al. and echoed in empirical studies, include:

• Domain-appropriate representation: Explanations must use the language and concepts familiar to stakeholders and avoid cognitive overload.
• Potential actionability: Identified features or patterns in explanations should be mappable to plausible interventions or follow-up actions.
• Consistency: Explanations should yield similar rationales for similar cases, ensuring fairness and avoiding disparate impact (Duell, 2021, Rojat et al., 2021, Sadeghi et al., 2023).

Regulations such as GDPR's "right to explanation" motivate transparent audit trails, while recent interactive frameworks (e.g. H-XAI) now enable users and auditors to test hypotheses and generate instance- or group-level fairness audits by combining classical XAI with causal effect estimation and bias assessments (Lakkaraju et al., 7 Aug 2025).

6. Current Challenges and Future Research

Open issues in XAI research include:

• Stability and faithfulness: Many methods (esp. LIME, SHAP) are susceptible to feature correlations and lack invariant attributions under perturbations. Advances in conditional attribution and causal modeling are needed (Salih et al., 2023, Islam et al., 2021).
• Evaluation and Sanity Checks: Metrics for faithfulness, robustness, and complexity are often baseline-dependent and sometimes rank dummy explanations above informative ones, underscoring the importance of multi-metric, sanity-checked evaluation (Stassin et al., 2023, Bommer et al., 2023).
• Local-global bridging: Current methods inadequately relate specific local explanations to global model logic. Hierarchical surrogates and clustering approaches (e.g., Global Attribution Mapping) are being explored (Sadeghi et al., 2023, Karim et al., 2022).
• Human-centered design: There is ongoing need for empirical user studies measuring how explanations affect trust, actionability, and error correction in domain workflows (Labarta et al., 14 Oct 2024, Duell, 2021).
• Semantic and causal explanations: There is momentum toward concept bottleneck models, integration with knowledge graphs, and the use of language and multimodal foundation models to generate high-level, semantically coherent explanations (Mumuni et al., 17 Jan 2025).

Emerging best-practice guidelines favor multi-method XAI pipelines, inclusion of domain knowledge, calibrated complexity, and iterative evaluation anchored in real end-user tasks. Advances in foundation-model-driven and causal grounding of explanations are expected to play major roles in the next generation of XAI systems (Mumuni et al., 17 Jan 2025, Lakkaraju et al., 7 Aug 2025).

7. Synthesis and Outlook

XAI encompasses a spectrum from interpretable modeling (GAMs, trees, rules) to sophisticated post-hoc explanations (SHAP, LIME, Anchors, Grad-CAM, IG, LRP) and new fusion pipelines leveraging large language and vision-LLMs for semantic, interactive, and multi-stakeholder explanations. Method selection and evaluation are inherently context-dependent; no single approach suits all scientific or operational challenges. The field is maturing toward principled, formally evaluated, and human-centered frameworks that balance model fidelity, stakeholder transparency, fairness, and real-world utility (Islam et al., 2021, Mumuni et al., 17 Jan 2025, Lakkaraju et al., 7 Aug 2025, Labarta et al., 14 Oct 2024).