PaperlayerView exact revision on arXiv

Source-linked AI summary

Explainable Artificial Intelligence: Precepts, Methods, and Opportunities for Research in Construction

Peter ED Love, Weili Fang, Jane Matthews, Stuart Porter, Hanbin Luo, Lieyun Ding

arXiv:2211.06579v2cs.AI

TL;DR

Construction has given limited attention to XAI, despite concerns about opaque AI outputs and their implications for confidence in decisions. The paper responds with a narrative review that develops a taxonomy of XAI precepts and methods, identifies future research opportunities, and notes that the review discourse remains evolving.

  • Problem

    Construction literature has given limited attention to XAI, while opaque AI outputs can undermine confidence in decisions.

  • Method

    The paper uses a narrative review to examine XAI and develops a taxonomy comprising its precepts and methods.

  • Results

    The review develops a taxonomy of XAI literature comprising precepts such as explainability, interpretability, and transparency and methods including transparent, opaque, and post-hoc approaches.

  • Takeaways & Limitations

    The review raises awareness about XAI in construction and identifies opportunities intended to stimulate a new research agenda.

  • Takeaways & Limitations

    The review is not systematic because XAI in construction is novel and only a limited number of studies have used it.

Abstract

from arXiv · show

Explainable artificial intelligence has received limited attention in construction despite its growing importance in various other industrial sectors. In this paper, we provide a narrative review of XAI to raise awareness about its potential in construction. Our review develops a taxonomy of the XAI literature comprising its precepts and approaches. Opportunities for future XAI research focusing on stakeholder desiderata and data and information fusion are identified and discussed. We hope the opportunities we suggest stimulate new lines of inquiry to help alleviate the scepticism and hesitancy toward AI adoption and integration in construction.

1.0 Introduction

The introduction frames XAI as a response to opaque AI models whose outputs can be difficult to interpret, creating risks for decision-makers and barriers to construction adoption. The paper addresses limited construction-focused XAI literature through a narrative review of its precepts, methods, and research opportunities.

  • AI models can uncover associations and patterns, but prior knowledge is needed to infer causal relationships.
  • DL and ML models are complex, nonlinear, and difficult for both laypeople and their creators to interpret or explain.
  • XAI makes AI behavior more intelligible by explaining how systems produce outputs so end-users can comprehend and trust them.
  • Construction reviews have largely overlooked XAI, leaving the literature relatively silent despite growing attention in other sectors.
  • Construction organizations have hesitated to embrace AI because autonomous decisions and actions are unexplainable, weakening confidence and trust in predictions.
  • The paper adopts a narrative review to examine XAI precepts and methods, develop a conceptual taxonomy, and identify future research opportunities in construction.

2.0 Explainable AI

This section distinguishes XAI precepts and organizes their relationships through a literature-derived taxonomy. It also highlights unresolved questions about trust and the meaning of explainability and interpretability.

  • Explainability and interpretability are often used interchangeably despite having distinct meanings in the XAI literature.
  • Explainability concerns clarifying a model’s internal mechanics in human terms, whereas interpretability concerns predicting effects of changes to inputs and algorithmic parameters.
  • Trustworthiness is associated with explainability and interpretability, but how trust is incorporated into existing ML models remains ambiguous.
  • Trust can be subjective, and model accuracy may be treated as sufficient for trust even when interpretability serves no purpose.
  • The literature-derived taxonomy maps relationships between XAI precepts and methods to structure the review.

2.1 Explainability

Explainability concerns how AI systems function and whether their outputs can be trusted, understood, and validated in construction. The paper frames XAI as a way to address limited questioning of construction AI outputs and to organize relevant concepts and stakeholder questions.

  • Concept and purpose: Explainability concerns the causes and functioning behind an AI system’s outputs, with understandable explanations shaped by the questions users ask.The paper contrasts explainability with the broader philosophical debate over what makes an explanation adequate.
  • Stakeholder needs: Stakeholder-oriented XAI questions include why a model was used, whether its output is correct, how recommendations can be trusted, and how errors can be corrected.The paper presents these questions as a contrast with the traditional AI focus in construction.
  • Scope of XAI: Explainability is treated as subjective because end-users judge whether an explanation is believable, while XAI also examines training data, activated features, and model layers.These dimensions include explainable data, explainable features, and explainable models.
  • Construction gap: Construction AI outputs are seldom questioned or validated by end-users in real-life contexts, especially for computer-vision applications.The usefulness of solutions addressing industry needs is also rarely examined in the academic construction literature.
  • Paper organization: The paper introduces XAI precepts and presents a conceptual taxonomy intended to organize the literature and approaches relevant to construction.The taxonomy is presented alongside a comparison of traditional AI and XAI in construction.

2.2 Interpretability

Interpretability is the passive, human-oriented property of a model being understandable through meaningful language or artifacts. The paper links interpretability to transparency, ethical decision-making, model structure, and the limits of simplified explanations.

  • Interpretability: Interpretability is a passive property based on a priori knowledge that enables a human observer to understand a model’s outcome.Its success depends on users’ cognition, knowledge, and biases.
  • Interpretability: Interpretability can be supported by visual or textual artifacts that make model behavior and outputs more transparent to users.The paper describes understandable language and readily understandable outcomes as central to this goal.
  • Implications: Transparency supports fair and ethical decision-making, but the appropriate balance between predictive performance and explanation depends on the model and context.The paper presents this balance as a recurring issue in applying AI to construction.
  • Transparency dimensions: Transparency is examined through simulatability, decomposability, and algorithmic transparency, covering whole-model behavior, components, and training procedures.These dimensions provide technical insights into how AI-based systems function.
  • Model complexity: Simple models can still become difficult to interpret when feature selection, preprocessing, engineered features, or model size obscure their components and relations.The paper notes that larger decision trees shift from simulatable to decomposable or algorithmically transparent forms.
  • Model complexity: Deep-learning models lack algorithmic transparency despite efficient heuristic optimization, and their performance on new similar problems is not guaranteed a priori.The paper uses construction cost prediction to illustrate how model behavior and reported error may remain unexplained.
  • Construction literature: Construction AI reviews have emphasized where systems are applied rather than what models require or how they function across the three transparency dimensions.The paper identifies this as a gap in prior construction-focused AI reviews.
  • Completeness: A central challenge is producing explanations that are both interpretable and complete, because simplified descriptions may hide limitations or conditions.The paper also notes that explanations can become persuasive rather than transparent when simplicity is preferred.

3.0 Explainability Methods

The paper organizes XAI methods into transparent models explained during training and opaque models explained after training. It also distinguishes methods that apply across models from methods tailored to particular model families.

  • Research gap: Construction research gives little attention to the trade-off between predictive performance and explainability.The paper identifies this trade-off as a recurring issue in the broader XAI literature.
  • Model types: XAI models are typically transparent or opaque: transparent models are explained during training, whereas opaque models are explained after training.Transparent models satisfy intrinsic transparency dimensions through their architecture; opaque models require post-hoc explainability.
  • Opaque models: Opaque models such as RF, SVM, CNN, MNN, and RNN are used when transparent methods provide unsatisfactory performance.Their black-box character motivates the use of post-hoc explainers.
  • Method scope: Model-agnostic explainability methods work across models, whereas model-specific methods are designed for particular models such as deep learning algorithms, SVM, and RF.The paper identifies these as distinct approaches within post-hoc explainability.
  • Review scope: The paper reviews transparent and opaque approaches to clarify how construction applications can address the black-box character of AI models.It presents the review as a way to examine methods used in construction rather than as an empirical comparison.

3.1 Transparent Models

Transparent models expose aspects of their structure or operation before or during prediction, supporting human understanding across several model families. Their transparency can diminish as models become larger, more complex, or combined with other techniques.

  • Model landscape: Transparent model families reviewed for construction include regression, decision trees, K-nearest neighbor, rule-based learners, and general additive models.The paper surveys these models as examples of approaches with intrinsic or structural explainability.
  • Regression models: Linear and logistic regression models are transparent structures whose explainability depends partly on whether users can understand their variables and outputs.Construction studies have reported high accuracy for these simple models.
  • Decision trees: Decision trees can be simulatable, decomposable, or algorithmically transparent depending on their size and the complexity of their feature relations.Small trees can be evaluated completely by users, while larger trees hinder complete evaluation and expose algorithmic procedures instead.
  • Decision trees: Decision trees have been used in construction for formwork selection and decision-support systems that classify roadway design options.Their off-the-shelf transparency supports these applications.
  • Ensemble methods: Combining decision trees into ensemble methods improves predictive performance but loses transparency and may require post-hoc explainability.Bagging and boosting are identified as common ensemble techniques in construction.
  • K-nearest neighbor: K-nearest-neighbor models classify observations through proximity and similarity, making their prediction rationale explainable through neighboring examples.Complex features or distance functions can restrict interpretability to the algorithmic operations themselves.
  • Rule-based learners: Rule-based learners represent knowledge through conditional or fuzzy rules that are generally understandable and transparent to users.Increasing rule number, length, coverage, or specificity can reduce interpretability even as system performance improves.

3.2 Opaque Models

Opaque models generally provide predictive accuracy but are difficult to interpret, so construction applications require post-hoc explanations and multiple techniques tailored to model type.

  • Opaque models generally outperform transparent ones in predictive accuracy, but their internal functioning is difficult to interpret.
  • Random Forest: Random Forest models improve single-tree accuracy by combining multiple trees, although the resulting ensemble is harder to interpret.
  • Post Hoc Explainability: Post-hoc methods—including simplification, local explanations, visualizations, ablation, permutation, random noise, and integrated gradients—help interpret opaque models without describing their exact operation.
  • Multi-layer Neural Network: Multi-layer neural networks require hyperparameter tuning through cross-validation and may suffer from vanishing gradients and questionable explainability.
  • Convolutional Neural Network: CNNs learn spatial feature hierarchies for vision tasks, but their complex internal relations make explanation difficult.
  • Recurrent Neural Network: RNNs model sequential data and long-term dependencies, yet their explainability remains challenging and their outputs require explicit caveats.

4.0 Opportunities for Research in Construction

The paper identifies stakeholder desiderata and data and information fusion as immediate opportunities for construction XAI research, alongside a need to move beyond algorithm-centered demonstrations toward practical usefulness.

  • Immediate construction XAI research opportunities center on stakeholder desiderata and data and information fusion.
  • Technical XAI advances should not be construction researchers’ immediate priority because computer science and engineering researchers are addressing algorithms and evaluation metrics.
  • Construction researchers should remain aware of emerging legal and regulatory developments even though developing requirements lies beyond their remit.
  • Construction studies often demonstrate that new DL or ML algorithms work on artificial problems but rarely establish their practical utility.
  • Future AI-enabled construction solutions should be based on stakeholder desiderata rather than only researchers’ enthusiasm for applying new algorithms.

4.1 Stakeholder Desiderata

Stakeholder-oriented XAI research should identify and manage differing desiderata, solicit users’ explanation needs, and develop domain-specific evaluation frameworks for understandable and usable systems.

  • Stakeholder desiderata include the differing goals, needs, interests, expectations, and demands of people who operate, improve, regulate, deploy, or are affected by AI systems.
  • Construction research must examine who manages and evaluates stakeholder desiderata and how those desiderata are solicited.
  • Construction managers need explanations that help them assess whether safety-risk outputs are accurate and reliable, potentially ranging from onboarding overviews to detailed reasoning.
  • End-users may ask why, how, performance, and training-data questions, and those questions can guide the choice of XAI technique.
  • Successful explanation processes require collaboration between domain experts and developers to translate XAI insights into understandable explanations for end-users.
  • Evaluation frameworks with measurable outcomes for domain-specific applications and stakeholder desiderata remain to be developed.

4.2 Data and Information Fusion

Data and information fusion is proposed as a construction XAI opportunity because combining sources may improve interpretability while introducing model-correlation, privacy, and explainability challenges.

  • For this paper, data fusion means assembling and combining different kinds of information into a procedure that yields a single model.
  • Construction studies, particularly computer-vision studies, should emphasize data fusion to improve interpretability and transparency explanations.
  • Fusion can occur at the raw-data, model-aggregation, or knowledge levels, with more complex systems requiring approaches such as federated and multi-view learning.
  • Deep-learning fusion couples data integration to the task and creates correlated features, producing a performance–explainability trade-off.
  • A multi-view perspective can clarify model behavior by examining how information from different single sources is fused.
  • Construction lacks empirical work on on-site data privacy, although privacy is identified as a critical barrier to applying XAI.

5.0 Conclusion

The conclusion argues that construction needs greater attention to XAI alongside predictive AI, because understanding model outputs supports more informed and confident decision-making. The review provides a taxonomy of XAI precepts and methods and identifies research opportunities centered on stakeholder desiderata and data/information fusion.

  • XAI can explain how DL and ML outputs are generated, supporting understanding of model behavior and outputs while reducing bias and error and improving confidence in decision-making.
  • Limited attention to XAI in construction contrasts with growing use of DL and ML to improve productivity and project and asset performance.
  • Understanding why and how DL and ML models produce predictions is important because opaque outputs can contribute to precarious decision-making and outcomes.
  • The review develops a taxonomy comprising XAI precepts and methods, including transparent and opaque models with post-hoc explainability.
  • The taxonomy is proposed as a reference for navigating the rapidly changing XAI literature.
  • Future research opportunities focus on stakeholder desiderata and data/information fusion, with accountability, fairness, discrimination, and unresolved legal ramifications requiring further consideration.
Loading 2211.06579v2…