A Comprehensive Taxonomy for Explainable Artificial Intelligence: A Systematic Survey of Surveys on Methods and Concepts

In plain words

Many people want to know why an AI system made a particular decision. The field that tries to answer this is called explainable AI. It has grown so fast that there are now many surveys, each using its own vocabulary and its own way of sorting methods. A newcomer, or a team choosing a method for a real product, has no common map.

The authors read a large number of these surveys, collected the words they use for the same ideas, and merged them into one structured list. The list starts with the problem: what task the model does, and how easy it is to understand. It then describes the explanation tool itself: what it takes in, what it outputs, how a person talks to it, and what rules it must follow. The last part covers how to judge whether an explanation is good, either with automatic measures or with people.

The paper is not an experiment and does not say which method is best. It is a map plus a reading list. It helps a beginner get oriented, helps a practitioner write down what the explanation must do, and helps a researcher see which combinations of properties have been little studied.

TL;DR

• A structured meta-study of the surveys on XAI methods, metrics and method traits; the search ended in early 2021 §3.3.
• Merges their terms into a taxonomy with three top-level parts: problem definition, explanator properties, and metrics §5.
• Illustrates each aspect with more than 50 example methods, summarised in one table tab. 6.
• No experiments and no numeric method results; use it to choose what to look for, then read the source surveys or methods.

Problem

Learned models are often black boxes, and laws, safety standards and business needs ask for explanations §1. Choosing a suitable XAI method requires a complete list of the properties that matter for a use case. The authors argue that earlier taxonomies each cover only part of this list, and that the surveys that contain them are scattered and hard to compare §1, §2.1. The paper does not cover use-case analysis or stakeholder requirements; it explicitly leaves them out §5.1.

Method

The paper has two parts: a structured review of the literature §3, §4, and a taxonomy built from it §5. The paper gives no algorithm. The sketch below only restates the selection procedure of §3.1 in code form; it is not the authors' code.

# Survey selection, restated from the paper's prose  [§3.1]
# Input:  search phrases; Google Scholar result lists
# Output: surveys kept for the survey of surveys

def select_surveys(search_phrases):
    # Scan the first 300 results of each search by title  [§3.1]
    candidates = []
    for phrase in search_phrases:
        candidates += first_300_results(phrase)

    # Keep promising titles, then check their abstracts  [§3.1]
    shortlist = [c for c in candidates if relevant_by_title(c)]
    kept = [c for c in shortlist if relevant_by_abstract(c)]

    # Add surveys found in reference lists or suggested by reviewers  [§3.1]
    kept += referenced_reviews(kept)

    # Drop surveys that are too narrow, too short or rarely cited  [§3.2]
    return [c for c in kept if not too_narrow_or_short_or_uncited(c)]

Literature review §3

1. Two search iterations. The first looks for work on XAI taxonomies using term combinations (AI, DNN, deep learning, ML; XAI, explain, interpret; taxonomy, framework, toolbox, guide) in Google Scholar §3.1 tab. 1. The second looks for general XAI surveys using "XAI" and "XAI survey" §3.1.
2. Each search is scanned by title for the first 300 results, then promising ones by abstract. Only works the authors could access are kept. Reference lists are scanned recursively for further reviews §3.1.
3. Surveys are sorted by general focus (general method collection, domain-specific collection, conceptual review, toolbox), length, target audience (beginner, practitioner, researcher), citations per year, and recency §3.2.
4. Surveys that do not fit the search target, are too narrow, are very short (up to 4 pages) or are rarely cited are dropped §3.2.
5. Over 70 surveys were reviewed; more than 50 are discussed in §4, along with some toolboxes §3.3 tab. 2 tab. 3 tab. 4 tab. 5.

Survey of surveys §4

6. Conceptual reviews are grouped as broad, stakeholder-perspective, HCI-perspective and evaluation-focused §4.1 tab. 2.
7. General method collections are listed by length and citations per year §4.2 tab. 3. Specialised collections (by domain, technology or method trait) follow §4.3 tab. 4. Toolboxes are listed §4.4 tab. 5.

Taxonomy §5. The root is procedural: first define the problem, then describe the explanator, then evaluate it §5 fig. 3.

8. Problem definition §5.1: fix the task (task type: clustering, regression, classification, detection, segmentation; input data type: tabular, text, graph, image, point cloud, audio) §5.1.1. Then fix the interpretability of the explained model §5.1.2 fig. 4:
   • intrinsically interpretable models, read as a whole (simulatable), decomposed into parts (decomposable), or known mathematically (algorithmic transparency) §5.1.2;
   • blended models that mix symbolic and sub-symbolic parts §5.1.2;
   • self-explaining models whose outputs include an explanation: attention maps, disentangled representations, or textual or multi-modal explanations §5.1.2;
   • post-hoc helper (surrogate or proxy) models that fully mimic the model or approximate one aspect of it §5.1.2.
9. Explanator §5.2 fig. 5: a function with an input, an output, interactivity and mathematical constraints.
   • Input §5.2.1: required inputs (model, data, user feedback, context); portability, from model-agnostic (only inputs and outputs are used) through hybrid (part of the internals) to model-specific §5.2.1; locality, local (valid near given samples) or global §5.2.1.
   • Output §5.2.2: object of explanation (processing, inner representation, development during training, uncertainty, data) §5.2.2; output type (example instance, contrastive or counterfactual, prototype, feature importance, rule, dimension reduction, dependence plot, graph, or a combination) §5.2.2; presentation (information units, accessibility, privacy awareness) §5.2.2.
   • Interactivity §5.2.3: the interaction task and the explanation process.
   • Mathematical constraints §5.2.4: linearity, monotonicity, satisfiability, number of iterations, size limits.
10. Metrics §5.3 fig. 6, divided by how much human judgement they need, after Doshi-Velez et al. §4.1:
    • functionally-grounded, no human judgement §5.3.1: faithfulness, localisation accuracy, completeness or coverage, overlap, surrogate accuracy, architectural and algorithmic complexity, stability, consistency, sensitivity, expressiveness §5.3.1;
    • human-grounded, subjective judgement with lay users §5.3.2: interpretability or comprehensibility, effectiveness, time efficiency, degree of understanding, amount of information §5.3.2;
    • application-grounded, a full human-AI system is needed §5.3.3: satisfaction, persuasiveness, improvement of human judgement or of the joint system's performance, automation capability, novelty §5.3.3.
11. Example methods are attached to each aspect in grey in the text and summarised in tab. 6. The examples are selected for diversity and recency, not as a full list §3.3 §5.

Hyperparameters

The review's settings, in place of model hyperparameters.

Results

Author-reported. The paper reports no experimental results. Its outputs are counts and a classification. Setup: Google Scholar searches, time window 2010 to 2021, reviews up to early 2021 §3. Baseline: not applicable. The paper compares its coverage with earlier meta-studies: one reviews 53 surveys §2.1, one covers 18 visual-XAI surveys §2.1. Most reviewed surveys are from 2019 to 2021 §3.3 fig. 2.

Paper vs code

No official code. The paper is a survey and implements no method. The only repositories it names are toolboxes it reviews, such as Skater, InterpretML, iNNvestigate, AIF360, alibi and others tab. 5; these are third-party and are not used here. The paper has no project page. A GitHub search was not run, because a search for a survey would return only other people's code.

Pitfalls

• The literature window ends in early 2021, while the arXiv version was updated in 2023. Work after that is absent from the survey list and from the taxonomy examples §3.3 §6.
• Counts differ between places: the abstract says "more than 50" surveys were identified and reviewed; §3.3 says over 70 were reviewed and more than 50 were detailed. Use the §3.3 numbers for coverage.
• The search is shaped by its terms and by Google Scholar. The authors acknowledge the bias towards the sub-fields they searched §6.
• Target audience and "practically suitable" labels are judgements by the authors §3.2. Treat them as a guide, not a fact.
• The taxonomy aspects overlap. For example, "hybrid" is also called eclectic or grey-box, and portability is also called translucency or transferability §5.2.1. Pick one name per requirement and check the source surveys.
• Intrinsic interpretability is a property of the model class, not a claim that a given model is understandable. The paper lists the classes it considers inherently transparent §5.1.2.
• The examples are illustrations, not a benchmark. The paper says the list is not complete §5. Do not rank methods from their placement in the taxonomy.
• Metrics are only classified, not defined with formulas or measured. Read the method papers for exact definitions §5.3 §4.1.
• The paper says it is not a chapter outline of XAI and not a survey of methods in itself §1. Do not use the taxonomy as a method catalogue without reading the cited sources.

Lineage

• Builds on: the metric categories of Doshi-Velez et al. (functionally-grounded, human-grounded, application-grounded) §4.1 §5.3; the transparency levels of Lipton, and the portability notion of Yao et al. §5.1.2 §5.2.1; the definitions in Bruckert et al. §2.3.
• Compared against: Vilone et al., a systematic review of 53 surveys §2.1; Chatzimparmpas et al., 18 visual-XAI surveys §2.1; Mueller et al., the DARPA report §2.1.
• Superseded by: not stated in the paper.