Computer science researchers have developed a hierarchical taxonomy and decision tree framework designed to help artificial intelligence developers systematically select appropriate interpretability techniques for machine learning systems. As artificial intelligence tools are increasingly deployed to execute complex human tasks and extract statistical correlations from large datasets to inform critical decisions, engineers face mounting pressure to explain how these computational models derive their outputs. While theoretical investigations into explainable artificial intelligence have expanded rapidly in recent years, practitioners in production settings frequently lack standardized, operational mechanisms to implement these concepts. To bridge the divide between abstract theoretical literature and practical software engineering, researchers Carter Cousineau and Rozita Dara conducted a two-phase meta-analysis examining dozens of academic survey papers and created structured decision trees that guide developers through model-specific and model-agnostic explanation methods.
The study centers on the challenge of interpretability, an essential concept within trustworthy artificial intelligence research that focuses on understanding the inner workings, reasoning, and decisions generated by algorithmic systems. Artificial intelligence models are fundamentally underpinned by machine learning algorithms, which are broadly divided into white-box and black-box architectures. White-box systems possess transparent internal logic that allows stakeholders to directly inspect algorithmic pathways, whereas black-box models obscure their inner mechanics behind layers of computational complexity. This lack of transparency presents substantial operational and societal hurdles, particularly when critical human decisions rely on automated outputs. Previous research highlighted by the authors indicates that opaque machine learning structures can undermine human trust, hinder user adoption, and complicate efforts to detect poor predictions, mitigate overfitting, or uncover algorithmic bias.
To investigate the landscape of existing solutions, the authors implemented a rigorous research design centered on the survey of surveys methodology. Phase one of the investigation comprised a comprehensive analysis of the research literature published between 2019 and 2023, scanning academic repositories including Elsevier Science Direct, Wiley, IEEE Xplore, SpringerLink, MDPI, Nature Machine Learning, and the ACM Digital Library. The literature search established specific inclusion and exclusion criteria, filtering peer-reviewed works using structured search strings combining terms for artificial intelligence, interpretability, and taxonomies. From an initial pool of 70 papers, the authors conducted iterative screening rounds involving two independent reviewers to verify inclusion decisions and mitigate selection bias, ultimately selecting 30 comprehensive survey papers for detailed analysis, complemented by targeted foundational literature spanning from 2006 to 2024.
Analysis of this literature corpus revealed that the rapid proliferation of interpretability publications since 2019 has inadvertently created confusion and inconsistency for practitioners attempting to navigate the field. Many earlier taxonomies, such as those establishing global terminology or categorizing post-hoc explainability, operated at high levels of theoretical abstraction that proved difficult to translate into concrete software development workflows. Furthermore, existing practical tools examined in prior studies were frequently characterized by practitioners as being either overly rigid or excessively flexible. To address these operational constraints, the authors synthesized the findings from their survey of surveys at a lower level of abstraction, isolating the concrete technical dimensions and operational trade-offs that software engineers must evaluate when integrating explainability into the design, development, and deployment phases of the artificial intelligence lifecycle.
The synthesized findings established an eight-category hierarchical taxonomy that systematically organizes the core components of algorithmic interpretability. The first category examines the underlying machine learning algorithm types, distinguishing between inherently interpretable transparent models and complex architectures. Categories two and three evaluate model decomposition, focusing on the variations in artificial intelligence outputs and the characteristics of features, recognizing that features understandable to human end users are essential for decision support. Category four addresses algorithmic transparency by delineating between local predictions that explain a single inference using a data subset and global predictions that convey high-level operational behavior across entire datasets. The fifth category classifies the size, volume, and structure of data sources, contrasting structured numerical or categorical tables with unstructured inputs such as text, audio, images, and sequence files.
The remaining taxonomy categories provide technical criteria for selecting and applying specific analytical tools. Category six distinguishes model-agnostic techniques, which evaluate system behavior externally without requiring internal access to algorithmic weights, from model-specific methods tailored exclusively to particular architectures. Category seven catalogs common interpretability techniques identified across the literature, including rule extractions, layer-wise relevance propagation, sensitivity analysis, image-based heat maps, permutation feature importance, partial dependence plots, surrogate models, and game-theoretic formulations such as Shapley additive explanations alongside local interpretable model-agnostic explanations. Category eight addresses stakeholder understanding, emphasizing that different audiences—from compliance regulators and domain experts to engineers and end users—require distinct explanatory formats, ranging from instance-level breakdowns to broad spatial visualizations and non-linear trend representations.
In the second phase of the study, the authors operationalized these eight taxonomy categories into a sequence of actionable decision trees constructed with root nodes, intermediate decision nodes, and terminal leaf nodes. To demonstrate practical implementation, the authors developed sample decision trees tailored to two distinct algorithmic paradigms: ensemble and random forest models representing more transparent architectures, and recurrent neural networks representing complex deep learning systems. In both structures, developers begin at an initial entry point and navigate through a series of diagnostic questions addressing feature sensitivity, global versus local scope, data scale, and desired output representations. By answering these sequential questions at each decision node, the developer progresses along a defined pathway toward terminal leaf nodes that recommend suitable interpretability techniques, such as surrogate modeling, partial dependence plotting, or layer-wise relevance propagation.
To evaluate the real-world utility and clarity of the framework, the authors conducted preliminary testing and validation with a purposive sample of 12 artificial intelligence developers and industry practitioners recruited from previous studies, professional networks, and Amazon Mechanical Turk. The validation utilized a mixed-methods approach combining closed-ended scoring with open-ended qualitative inquiries, followed by thematic analysis using grounded theory methods. All 12 participating developers (100 percent) affirmed the value of the structured taxonomy and decision trees, reporting that the framework established a methodical foundation that mitigated ad hoc decision-making and helped narrow down an overwhelming volume of techniques into a focused subset aligned with specific project objectives and stakeholder needs.
Participant feedback also highlighted key operational tensions that engineers routinely encounter when implementing explainability in production environments. Developers emphasized that interpretability decisions cannot be made in isolation from computational constraints, noting that sophisticated post-hoc techniques often demand extensive execution time or resources. Furthermore, participants described an ongoing negotiation between transparency, predictive performance, and ethical obligations such as bias detection and fairness auditing. Because transparent architectures can sometimes achieve lower predictive accuracy than complex black-box networks, developers indicated that future iterations of the tool should provide explicit guidance on prioritizing overlapping taxonomy categories, navigating efficiency trade-offs, and establishing iterative feedback loops as machine learning systems evolve over time.
The authors noted specific methodological boundaries in their study, explaining that certain emerging model architectures—including Bayesian algorithms, generative adversarial networks, reinforcement learning systems, hybrid approaches, and large language models—were deliberately excluded from the initial taxonomy due to their inconsistent representation across meta-surveys published between 2019 and 2023. Additionally, while the preliminary testing confirmed conceptual clarity among a focused group of 12 practitioners, the authors stressed that future research must conduct larger, domain-specific empirical evaluations across fields such as finance and healthcare to measure long-term usability and assess how these developer tools ultimately affect end-user trust and decision outcomes.
Subject of Research: Artificial intelligence interpretability technique selection taxonomy and decision tree framework
Article Title: Interpretability taxonomy: an approach for artificial intelligence developers technique selection
Article References: Cousineau, C., & Dara, R. (2025). Interpretability taxonomy: an approach for artificial intelligence developers technique selection. Discover Data, 3(1), Article 60. https://doi.org/10.1007/s44248-025-00096-6
Image Credits: AI Generated
DOI: 10.1007/s44248-025-00096-6
Keywords: Artificial Intelligence, Machine Learning, Interpretability, Explainable AI, Decision Trees, Software Development, Trustworthy AI, Model Transparency, taxonomy, approach, artificial, intelligence
News Source: Blake Davidson. (November 7, 2026). Researchers Propose Decision Tree Framework to Help Developers Select AI Interpretability Techniques. Scienmag.
