Knowledge Graphs and Explainable AI as Complementar... | AI Research

Knowledge Graphs and Explainable AI as Complementary Resources for Urban Mining
This paper addresses the challenge of performing pre-demolition assessments (PDA) in urban mining, where AI tools must support human auditors who are ultimately responsible for regulatory compliance. The authors argue that simply having accurate AI predictions is not enough; the decisions supported by AI must be "defensible"—meaning they are legible, plausible, sourced, and contestable. The paper proposes a framework that integrates Explainable AI (XAI) with domain-specific Knowledge Graphs (KG) to create audit-ready documentation that neither technology could produce on its own.

The Need for Defensible AI

In urban mining, auditors must inventory building components and determine their future, such as reuse or recycling, in accordance with strict regulations. While XAI techniques like feature attribution can explain why a model made a specific prediction, these explanations often exist in a technical language that does not match the regulatory vocabulary required for audits. Conversely, Knowledge Graphs provide a structured, domain-specific map of building data but lack the ability to explain specific, real-time AI predictions. By combining these, the authors aim to create "accountability-bearing" audit artefacts that bridge the gap between raw AI output and professional regulatory requirements.

Four Modes of Integration

The authors define four specific ways to combine XAI and Knowledge Graphs, each unlocking a unique property of defensibility:

• Lifting: Translates raw AI feature attributions into the regulatory language of the Knowledge Graph, making the explanation "legible" to the auditor.

• Constraining: Filters AI-generated counterfactuals (hypothetical scenarios) through domain rules to ensure that any proposed intervention is "plausible" and physically possible within the building’s context.

• Typing: Attaches reliability scores and evidence chains to assertions, providing "sourcing" so the auditor knows exactly how much confidence to place in a specific piece of data.

• Revising: Creates a record of auditor actions, such as contesting or correcting a claim, which ensures the decision process is "contestable" and maintains a clear history of how conclusions were reached.

Practical Application

To demonstrate these concepts, the authors use the example of a fire door during a building assessment. In this scenario, the AI might identify a component as a "fire door." Through the Lifting mode, this is mapped to specific building components like seals and hardware. If the auditor questions the reuse value, the Constraining mode filters out impossible suggestions and highlights actionable changes—such as using bolted hinges instead of glued ones—that would actually improve the door's recyclability. Finally, the Typing and Revising modes allow the auditor to verify the evidence for the door's fire rating and correct the record if they discover the door is actually for smoke control, ensuring the audit trail remains accurate and transparent.

Considerations for Implementation

The authors emphasize that this framework is a structural guide rather than a set of empirical results. It provides a vocabulary for how to build AI systems that are "defensible by construction." A key limitation noted is that the framework does not replace human authority; while it tracks revisions and provides evidence, the final responsibility for validating these changes under a governance regime remains with the human auditor. The authors suggest that this approach is particularly relevant for meeting emerging regulatory standards, such as the EU AI Act, which increasingly require documented, transparent reasoning for automated decisions in high-stakes environments.