Earticle

초록

영어

Explainable artificial intelligence (XAI) has expanded from a niche concern into a defining requirement for the deployment of opaque machine-learning models, yet the field's rapid growth has produced a fragmented body of evidence in which the same handful of techniques are applied across very different domains with little agreement on how explanations should be evaluated. This study maps the structure of that evidence rather than pooling its effect sizes. Departing methodologically from variance-weighted metaanalysis, we conduct a systematic mapping with structured content-analytic coding and bibliometric network synthesis over a tri-lingual corpus spanning the English, Chinese, and Korean literatures from 2021 to 2025 (coded sample N = 240; macro indicators triangulated against an indexed corpus of 4,781 articles). Each study is coded for application domain, XAI method, explanation scope and locality, and evaluation practice, and the resulting landscape is analysed through method-prevalence profiles, a method×domain co-occurrence matrix, a method co-occurrence network, and a five-year compositional trend. Three findings stand out. First, the field is strongly SHAP-centric: post-hoc, model-agnostic attribution dominates, while intrinsically interpretable glassbox models remain marginal. Second, method choice is largely dictated by data modality rather than by explanatory goal, producing a stable but narrow method×domain coupling. Third, and most consequentially, a persistent evaluation gap remains: a plurality of applied studies present explanations illustratively, without faithfulness, robustness, or human-grounded validation. We argue that the central problem facing XAI is not a shortage of methods but a shortage of disciplined evaluation, and we outline an evaluation-first agenda and a minimal reporting standard for explainability research.

목차

Abstract
1. INTRODUCTION
2. CONCEPTUAL BACKGROUND
2.1. From Interpretability to Explainability
2.2. A Taxonomy of XAI Methods
2.3. Why Mapping Rather Than Meta-Analysis
3. RESEARCH DESIGN
3.1. Search Strategy and Corpus Construction
3.2. Coding Scheme
3.3. Analytical Procedure and Principles of Cautious Inference
4. RESULTS
4.1. Corpus Overview: Accelerating, Tri-Lingual, and Health-Concentrated
4.2. Method Prevalence: A SHAP-Centric Field
4.3. Method × Domain Structure: Modality Dictates Method
4.4. Method Co-Occurrence: A Dense Post-Hoc Core
4.5. The Evaluation Gap
4.6. Temporal Evolution of Method Families
4.7. Cross-Language Contrasts
5. DISCUSSION
5.1. Core Findings
5.2. Theoretical Implications: Toward Evaluation-First Explainability
5.3. Practical Implications
5.4. Methodological Implication: A Minimal Reporting Standard for XAI
6. LIMITATIONS AND FUTURE RESEARCH
7. CONCLUSION
REFERENCES

키워드

저자

• Yituo Feng [ Information Management and Information Systems, Lancang–Mekong International Vocational Institute, Yunnan Minzu University, Kunming, China ]
• Qiufen Xu [ School of Economics and Management，Yunnan Minzu University, Kunming, China ]
• Dongshou Fan [ School of Economics and Management，Yunnan Minzu University, Kunming, China ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  Journal of Digital Intelligence and Management Systems

• 간기

  연3회

• eISSN

  3140-5851

• 수록기간

  2026~2026

• 십진분류

  KDC 569 DDC 620