Explainable Artificial Intelligence Approaches: A Survey

1 INTRODUCTION

• AI decisions affect human interests, rights, and lives, making explainability important in high-stakes settings such as credit approval, defense, and cybersecurity.The introduction also notes regulations addressing fairness, accountability, and transparency risks.Source: §1 · p. 1
• XAI has renewed relevance amid expanded AI use and concerns about unethical use, opacity, and bias; black-box decisions also remain problematic in industrial control.The paper describes explainability as broader than interpretability in importance, completeness, and fidelity of prediction.Source: §1 · p. 1
• Comparing XAI methods is difficult because approaches vary across global versus local, ante-hoc versus post-hoc, and surrogate explanations, while comprehensive analysis across these perspectives remains lacking.The introduction defines global explanations as covering the dataset and local explanations as addressing an individual instance.Source: §1 · pp. 1–2
• Explainability evaluation, especially evaluation based on human studies, remains under-focused despite progress in XAI methods.The introduction lists feature importance, interactions, Shapley values, partial dependence, surrogates, counterfactuals, adversarial methods, prototypes, and knowledge infusion among common concepts.Source: §1 · p. 2
• The survey demonstrates popular XAI methods on credit-default prediction, compares their trade-offs across perspectives, discusses explainability quantification, and recommends directions toward responsible or human-centered AI.It presents its mutual test case as a feature distinguishing its survey from other recent reviews.Source: §1 · p. 2

2 BACKGROUND

• XAI encompasses ML or AI systems for demystifying model internals (e.g., what models have learned) and/or explaining individual predictions [6] and.Source: §2 · p. 2
• Mueller et al. [9] characterize explanation systems in three generations: expert systems, human-computer systems designed for cognitive support, and recent tools clarifying mostly black-box systems.The third generation begins with the recent renaissance starting from 2015.Source: §2 · p. 2
• Three notions of explainability quantification include two types of human experiments testing whether people can predict model outcomes and a human-free proxy-task approach.A proxy example uses decision-tree depth as an indicator of explainability.Source: §2 · p. 2
• Prior reviews cover some perspectives, but a comprehensive review combining the important perspectives with a mutual test case is still missing.The paper proposes an overview using a shared task, multi-perspective analysis, and future directions toward responsible or human-centered AI.Source: §2 · p. 2

3 TEST CASE

• The shared test case predicts whether a customer will default on a mortgage payment in the near future and supports comparison of XAI approaches.The authors describe the task as helping explain the comparative advantages of different approaches.Source: §3 · pp. 2–3
• The experiments use the popular Freddie Mac dataset, with selected features and descriptions drawn from its user guide.Source: §3 · p. 3
• The authors use the R package iml to produce results for the XAI methods reviewed.Source: §3 · p. 3

4 EXPLAINABLE ARTIFICIAL INTELLIGENCE METHODS

• The section summarizes XAI methods through their advantages, disadvantages, challenges, and competitive strengths, drawing primarily on two comprehensive surveys and [16].Source: §4 · p. 3
• The survey enhances earlier reviews with multi-perspective analysis, recent research progress, and future research directions.Source: §4 · p. 3
• One cited taxonomy groups explanation methods as intrinsically interpretable methods, model-agnostic methods, and example-based explanations [16].Source: §4 · p. 3

4.1 Intrinsically Interpretable Methods

• Linear regression explains predictions through feature weights, but correlated features can make their distinct influences indeterminable.The coefficient explanation is most useful when the number of features is small.Source: §4.1 · pp. 3–4
• Logistic regression models class probabilities, while its weights indicate direction and degree of influence rather than additive contributions to the prediction.Source: §4.1 · p. 3
• Decision trees split data by thresholds; root-to-leaf paths explain decisions, and cutoffs can express counterfactuals even with nonlinear relationships or interacting features.The passage gives a credit-score cutoff of 748 as an example and notes that shorter trees are more explainable.Source: §4.1 · p. 3
• Tree explanations cannot express linear input-output relationships, can be sensitive to slight input changes, and become harder to interpret as trees grow deeper.Different trees can also represent the same problem.Source: §4.1 · p. 3
• Decision rules use simple IF-THEN conditions but are mostly limited to classification and inadequate for describing linear relationships.RuleFit learns sparse linear models and decision rules to detect interaction effects.Source: §4.1 · p. 3
• Linear-model extensions, Naïve Bayes, and K-Nearest Neighbors are also intrinsically interpretable, though extensions add complexity and natural explainability can trade off against performance.GLMs and GAMs address some linear-model assumptions but become less interpretable; Naïve Bayes assumes strong feature independence, while K-Nearest Neighbors predicts from nearby data points.Source: §4.1 · pp. 3–4

4.2 Model-Agnostic Methods

• Model-agnostic methods separate explanation from the model, allowing compatibility across models and different forms of explanation, including plots and linear formulas.The listed methods include PDP, ICE, ALE, feature interaction, feature importance, global and local surrogates, and SHAP.Source: §4.2 · p. 4

4.2.1 Partial Dependence Plot (PDP)

• PDP shows global marginal effects of one or two features on predictions, but assumes plotted features are uncorrelated with the rest, an unusual condition in real-world data.It averages effects across instances and has a practical limit of two features that it can clearly explain at once.Source: §4.2.1 · p. 4, Figure 2 · p. 4

4.2.2 Individual Conditional Expectation (ICE)

• ICE plots one prediction-effect line per instance, whose average gives the PDP, but many lines can overcrowd the plot and correlated features remain problematic.ICE can meaningfully display only one feature at a time.Source: §4.2.2 · p. 4, Figure 3 · p. 5, Figure 4 · p. 5

4.2.3 Accumulated Local Effects (ALE) Plot

• ALE plots describe average feature effects, work reasonably well with correlated features, and are comparatively faster than PDP, but strongly correlated features are hard to interpret in isolation.For strongly correlated features, the text says plots showing changes in both features together make more sense.Source: §4.2.3 · pp. 4–5, Figure 5 · p. 5

4.2.4 Feature Interaction

• Feature-interaction analysis uses Friedman’s H-statistic to detect interactions, which make individual feature effects non-additive; current Actual UPB has the highest interaction and credit score the least.Calculating interactions is computationally expensive, and sampling rather than using the full dataset can produce run-to-run variance.Source: §4.2.4 · p. 5

4.2.5 Feature Importance

• Current Actual UPB has the highest feature importance and interaction strength (6), while credit score ranks lowest; interactions and run-to-run variance complicate interpretation.The account also notes uncertainty about using training versus test data and says importance does not sum to the total performance drop when features interact.Source: §4.2.5 · p. 6, Figure 7 · p. 6

4.2.6 Global Surrogate

• A global surrogate approximates a black box with an interpretable model, but multiple possible trees can yield different explanations and may create only an illusion of interpretability.The example approximates a Random Forest with CART decision trees.Source: §4.2.6 · p. 6, Figure 8 · p. 6

4.2.7 Local Surrogate (LIME)

• LIME explains individual predictions with a local interpretable model, but local fidelity does not ensure global approximation or stable explanations for nearby instances.In the example, the black box predicts default at probability 1 and LIME at 0.99; incomplete attribution may also make it unsuitable where complete legal explanations are required.Source: §4.2.7 · p. 6, Figure 9 · p. 7

4.2.8 Shapley Values

• Shapley values fairly allocate prediction contributions across feature coalitions and provide full explanations, but require 2^k coalitions and remain vulnerable to correlated features.For the sample, prediction 1.00 differs from the dataset average 0.53 by 0.47, with Current Actual UPB contributing 0.36.Source: §4.2.8 · p. 7

4.2.9 Break Down

• Break Down greedily adds features by contribution to the baseline; it is faster than Shapley and matches it without interactions, but contributions depend on prior additions.The figure visualizes positive or negative contributions from participating features toward one sample’s final prediction.Source: §4.2.9 · p. 7

4.3 Example-Based Explanations

• Example-based explanations use selected dataset instances to explain model behavior and data distribution, framing a prediction through similarity to cases and their outcomes.The section introduces counterfactuals, prototypes, influential instances, and nearest-neighbor explanations as examples.Source: §4.3 · p. 7

4.3.1 Counterfactual

• Counterfactuals describe input changes that could reverse an individual prediction, but multiple valid explanations complicate selection and high-cardinality categorical variables are difficult.The method does not require access to data or models; the credit example says changing a score from 748 to 749 reverses the predicted default with other features unchanged.Source: §4.3.1 · p. 7

4.3.2 Adversarial

• Adversarial examples use small intentional input perturbations to flip predictions, exposing vulnerabilities; in the credit example, raising a score by one reverses the decision.The passage notes threats to cybersecurity and autonomous vehicles, while adversarial examples may also help discover vulnerabilities and improve models.Source: §4.3.2 · pp. 7–8, Table 3 · p. 8

4.3.3 Prototypes

• Prototypes represent data well and criticisms represent it poorly, but choosing their optimal number is challenging; nearby customers 5 and 6 may switch classes across runs or models.Customers 1 and 10 are given as prototype examples.Source: §4.3.3 · p. 8, Figure 11 · p. 9

4.3.4 Influential Instances

• Influential training instances affect predictions and parameter determination and can aid debugging, but setting the cutoff between influential and non-influential cases is difficult.Customers 1, 2, 9, and 10 are examples of influential instances; customers 5 and 6 lie near the decision boundary and are not influential.Source: §4.3.4 · p. 9

4.3.5 k-nearest Neighbors Model

• A nearest-neighbor explanation shows which nearby samples support a prediction; the analysis identifies a gap for methods that are actual, direct, model-agnostic, and local.The example classifies a new sample using three neighbors, one gray and two white.Source: §4.3.5 · pp. 9–10, Figure 12 · p. 9

4.4 Other Techniques

• Instance-wise feature selection extracts the most informative features for a particular instance by maximizing mutual information with response variables, but is mostly limited to post-hoc approaches.Source: §4.4 · p. 10
• Conditional mutual information can quantify an explanation relative to a recipient’s background and personalize it, but the described approach does not comprehensively cover varied user needs.The example contrasts recipients who know linear algebra with those who do not.Source: §4.4 · p. 10
• Gradient-based attribution assigns contribution or relevance to input features for each output neuron, with heatmaps distinguishing features that support or suppress target activation.Source: §4.4 · p. 10
• A CNN visualization survey categorizes work by intermediate representations, feature-space diagnosis, filter-pattern disentanglement, interpretable CNNs, and semantic disentanglement; layer-wise relevance propagation maps abnormalities to features in industrial control systems.Source: §4.4 · p. 10

4.5 Knowledge Infusion Techniques

• Testing with CAV quantifies a prediction’s sensitivity to a user-defined concept, but the described approach is post-hoc and limited to image classification.The example measures the influence of “striped” on a “zebra” prediction as a single number.Source: §4.5 · p. 10
• Knowledge graphs can measure information loss in neural-network latent features and provide external knowledge to supervise feature learning.The cited work anticipates a role for knowledge graphs in explainable AI systems.Source: §4.5 · p. 10
• Infusing domain principles such as the 5C’s of credit or CIA principles reportedly improves explainability with negligible performance compromises, faster execution, and better generalization to unknown samples.Source: §4.5 · p. 10
• Domain-principle approaches combine model agnosticism, local and global explanations, and authenticity, but are not fully off-the-shelf because they require domain-specific configuration.The passage notes that further work remains.Source: §4.5 · p. 10

5.1 Quantifying Explainability

• Explainability evaluation remains an open challenge, approached through model-complexity measures and human studies; Table 4 compares methods across five key perspectives.Source: §5.1 · p. 10, Table 4 · p. 11

5.1.1 Model Complexity-based Explainability Evaluation

• Model-complexity measures include rule-list size, rule quality, interactions, and cognitive chunks,,, [58]; reducing tree size can raise error and feature selection is costly.Source: §5.1.1 · p. 11

5.1.2 Human Study-based Explainability Evaluation

• Human-study evidence spans task/person-relative interpretability [24], decision-table advantages [26], transparent small-feature models’ simulation and trust [27], and operation-count effects [25].Comprehensive human studies and quantifying human comprehensibility remain open challenges.Source: §5.1.2 · pp. 11–12

5.2 Future Research Directions

• The proposed research agenda aims at human-centered AI in which agents collaborate with people interpretably and explainably to support fairness, transparency, and accountability.Source: §5.2 · p. 12

5.2.1 A Generic Framework to Formalize Explainable Artificial Intelligence

• A generic XAI framework could address missing formalism, recipient-specific explanations, quantification, and human comprehensibility; prior domain-knowledge infusion work reported competitive performance with enhanced explainability, [51].Source: §5.2.1 · p. 12

5.2.2 Towards Fair, Accountable, and Transparent AI-based Models

• Responsible AI requires bias remediation, accountability to affected parties, and comprehensible transparency; explanations can uncover potential fairness risks and increase understanding and trust.Source: §5.2.2 · p. 12

5.2.3 Human-Machine Teaming

• Human-machine teaming calls for follow-up interaction and context-aware explanations, with human feedback and HCI empirical work proposed as routes toward adaptive, human-centered AI.Source: §5.2.3 · p. 12

5.2.4 Collective Intelligence from Multiple Disciplines

• The authors call for multidisciplinary research combining explanation studies, social and behavioral sciences, HCI, and knowledge from domains including healthcare, finance, medicine, security, and defense.Source: §5.2.4 · pp. 12–13

• Post-hoc explanations can mislead despite plausible appearance; for high-stakes decisions, cited authors recommend simple intrinsically explainable models, while pre-modeling is viable but under-focused.Source: §6 · p. 13
• Domain-knowledge infusion may enhance explainability but remains under-focused, prompting a call for more research on explaining black-box models with domain knowledge.Source: §6 · p. 13
• The conclusion calls for evaluating or quantifying explainability with both human and non-human studies.Source: §6 · p. 13
• The survey demonstrates and analyzes XAI methods using a shared test case to compare advantages and identify challenges and future directions.Source: §6 · p. 13