🤖 AI

Knowledge Graphs and Explainable AI as Complementary Resources for Urban Mining

This paper proposes a complementarity-theoretic framework for integrating Knowledge Graphs and Explainable AI in urban mining pre-demolition assessments, defining four specific modes (Lifting, Constraining, Typing, and Revising) that enhance the defensibility and regulatory compliance of auditor decisions beyond mere prediction accuracy.

Original authors: Jan Gronewald, Andreas Emrich, Nijat Mehdiyev

Published 2026-07-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jan Gronewald, Andreas Emrich, Nijat Mehdiyev

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: What is this old building made of, and can we save its parts for reuse? This is the job of "Urban Mining." But there's a catch: you can't just guess. You have to be a qualified auditor, and every decision you make must be able to stand up in court (or at least in front of a strict regulator).

Enter the AI. It's like a super-smart, super-fast assistant who can look at a building and say, "That's a fire door!" But here's the problem: the AI is a bit of a black box. It gives you an answer, but it doesn't always explain why in a way that fits your official paperwork. If the AI says, "I'm 91% sure," but you don't know what it looked at to get that number, you can't sign off on it.

This paper suggests a clever way to fix that. Instead of just trusting the AI or just trusting a human, the authors propose teaming them up using two special tools: Explainable AI (XAI) and Knowledge Graphs (KG). Think of XAI as the AI's "inner monologue" (why it thinks what it thinks) and the Knowledge Graph as a giant, organized library of rules about how buildings actually work.

The authors suggest that when you combine these two, you get four superpowers that make your decisions "defensible" (meaning you can prove they are right). They call these four modes Lifting, Constraining, Typing, and Revising. Let's break them down with a story about a Fire Door.

1. Lifting: Translating "Pixel Speak" into "Human Speak"

Imagine the AI looks at a photo of a door and sees three bright spots (highlights) that made it think, "Fire Door!" In the AI's world, these are just numbers and pixels.

The Problem: You can't write "Pixel 45 is bright" in your official report. You need to say "The door seal is intact."
The Solution (Lifting): The Knowledge Graph acts like a translator. It takes those bright spots and says, "Ah, that bright spot is actually the door seal, and that one is the door leaf."
The Result: The AI's explanation is "lifted" from confusing math into clear, official language. Now, your report says exactly what the AI saw, using the right words.

2. Constraining: The "Reality Check" Filter

Now, imagine you ask the AI, "What if we changed this door to make it worth more money?" The AI might suggest, "Swap the fire-proof core for a cheap plastic one!"

The Problem: That sounds great for profit, but it's illegal! You can't put a plastic core in a fire wall.
The Solution (Constraining): The Knowledge Graph is like a strict bouncer at a club. It has a list of rules (axioms) about what is allowed. It looks at the AI's wild idea and says, "Nope. That door is in a fire wall, so it must have a fire-rated core. That idea is impossible."
The Result: The AI only suggests changes that are actually possible and legal. It filters out the "magic" ideas that break the laws of physics or building codes.

3. Typing: The "Trust Score" Tag

The AI says, "I think this is a fire door." But how sure is it? Is it a wild guess or a solid fact?

The Problem: If you just write "Fire Door" in your report, no one knows if you should double-check it or if you can just move on.
The Solution (Typing): The system attaches a "Trust Tag" to every piece of info.
- If the AI saw it clearly, the tag says "Confirmed."
- If the AI is just guessing based on patterns, the tag says "Assumed" (with a score like 0.62).
- If the AI is totally unsure, it might say "Deleted."
The Result: You know exactly how much to trust each part of the report. If a tag says "Assumed," you know to do a quick visual check. If it says "Confirmed," you can move on.

4. Revising: The "Undo Button" with a Paper Trail

Imagine you do a quick check and realize, "Wait, this isn't a fire door; it's just a smoke-control door!"

The Problem: If you just erase the old note and write a new one, nobody knows who changed it or why. Later, someone might ask, "Why did you change your mind?"
The Solution (Revising): The system doesn't just delete the old info. It creates a history log. It records: "Auditor Jane changed 'Fire Door' to 'Smoke Door' at 2:00 PM because she saw the seal wasn't intumescent."
The Result: Your decision is now a living document. You can look back and see exactly how the story changed. If you made a mistake, it's recorded, but it's also fixed and explained.

The Big Picture

The authors aren't saying this is a magic wand that solves everything instantly. They are suggesting that by itself, AI isn't enough for these serious jobs. An AI without a rulebook (Knowledge Graph) might give you a cool explanation that makes no sense in the real world. A rulebook without an AI is just a static list that can't react to new buildings.

But when you combine them using these four steps, you get a system where the AI does the heavy lifting, and the human auditor stays in charge, with a clear, legal, and traceable reason for every single decision.

The paper uses a specific example of a wooden door in an office building to show how this works. The AI spots the door, the system translates the pixels, checks the rules, assigns a trust score, and lets the auditor correct a mistake while keeping a perfect record of the change.

The authors admit this is a positional paper—it's a strong argument and a new way of thinking, not a final proof that this works in every single building yet. They suggest that future work will need to test this in real-world audits to see if it really helps auditors do their jobs better. But the idea is solid: if you want AI to help with important, regulated decisions, you need to build a bridge between the AI's brain and the human's rulebook.

Technical Summary: Knowledge Graphs and Explainable AI as Complementary Resources for Urban Mining

Problem Statement
Urban mining, specifically the pre-demolition assessment (PDA) of existing building stock, relies on a regulated audit process where qualified auditors must inventory components and determine their fate (reuse, recycling, or disposal). While Artificial Intelligence (AI) can support these auditors, the primary constraint on scaling urban mining is not the prediction accuracy of technical models, but the defensibility of the decisions they support. Decisions must be legible in regulatory vocabulary, plausible under domain constraints, traceable in documentation, and contestable.

Current Explainable AI (XAI) techniques (e.g., feature attribution, counterfactuals) operate on the input space of black-box models, producing outputs that often lack alignment with regulatory categories or physical domain constraints. Conversely, Knowledge Graphs (KGs) provide structured domain knowledge but lack responsiveness to specific model instances. Existing literature has taxonomized the integration of KGs and XAI but remains structurally under-specified; it lacks a theoretical account of why specific integrations produce artifacts that neither resource can provide alone.

Methodology and Theoretical Lens
The paper adopts a complementarity-theoretic interpretation grounded in the Information Systems (IS) resource-based tradition. In this view, two resources are complementary when the marginal contribution of one increases with the level of the other. The authors argue that the integration of KGs and XAI is structural rather than aggregative: the joint object possesses properties (legibility, plausibility, sourcing, contestability) that the unilateral objects lack.

The framework operates on the "explanatory surface" of black-box predictors. It defines a property-state graph $\Sigma(x)$ for an instance $x$ , consisting of tuples $(e, p, v, r)$ representing entities, properties, values, and reliability annotations. The methodology proposes four consolidated integration modes, each defined as a typed operation over XAI artifacts and KG substrate structures:

Lifting: Maps XAI attributions (e.g., SHAP values) to KG instances. It uses a grounding function $\gamma$ $γ$ to map input features to typed instances (e.g., a salient region to a product:DoorLeaf) and aggregates salience masses along a mereological hierarchy.
- Defensibility Property: Legibility (explanations exist in the regulatory vocabulary at the required granularity).
Constraining: Filters XAI counterfactual candidates through a plausibility predicate based on KG compatibility axioms. It ensures that proposed interventions (e.g., changing a material) satisfy definitional dependencies and structural constraints (e.g., fire-compartment topology).
- Defensibility Property: Plausibility (counterfactuals correspond to interventions admissible for real audit decisions).
Typing: Constructs reliability annotations for property assertions based on evidence chains (model confidence, inference chains, calibration data). It applies a reliability typology (e.g., Assumed, Confirmed, Deleted) to specific assertions.
- Defensibility Property: Sourcing (differentiating the auditor's response based on the evidence supporting each property).
Revising: A bidirectional operator where auditor actions (accepting, contesting, or adding claims) update the property-state graph and generate provenance trails (using PROV-O). Cross-instance actions feed an evidence pool for ontology revision.
- Defensibility Property: Contestability (the audit decision is an inspectable, revisable artifact with a queryable history).

Key Contributions
The paper makes three primary contributions:

Theoretical Interpretation: It provides a complementarity-theoretic account of KG–XAI integration for accountability-bearing expert processes, moving beyond descriptive taxonomies to a structural analysis of how technical pairings become audit artifacts.
Consolidated Modes: It formalizes four specific integration modes (Lifting, Constraining, Typing, Revising), linking each to a distinct defensibility property required for regulatory compliance.
Illustrative Framework: It demonstrates the framework using a fire-door example within the context of the W3C Linked Building Data (LBD) stack and the Urban Mining Index.

Results and Illustrative Example
The paper does not present empirical experimental results but offers a worked example of a fire door assessment to illustrate the modes:

Lifting translates model saliency (seal, leaf, hardware) into a structured description of the door components, allowing the auditor to copy the explanation directly into audit fields.
Constraining filters counterfactuals; for instance, it rejects a suggestion to change the core material to a non-fire-rated substance because the door is in a fire-compartment wall (a constraint encoded in the KG), but accepts a change from glued to bolted hinges as a valid intervention.
Typing assigns reliability scores (e.g., Assumed with a confidence of 0.68) to the fire-resistance class, prompting the auditor to perform a specific clarifying inspection rather than a generic re-photograph.
Revising allows the auditor to contest the classification after inspection (finding the door is smoke-control, not fire-rated). This updates the property state, records the provenance of the change, and flags the ontology for potential refinement.

The authors note that removing any single mode degrades the defensibility profile: removing Lifting leaves only heatmaps; removing Constraining allows regulatorily forbidden substitutions; removing Typing leaves unsourced confidence scores; and removing Revising results in static inventories with no provenance trails.

Significance and Claims
The paper positions itself as a positional rather than empirical work. Its significance lies in offering a structural vocabulary for integrating KGs and XAI in expert-domain AI where a qualified human remains accountable.

The authors claim that cautionary findings regarding XAI in expert domains (e.g., explanations failing to improve accuracy) may be structurally explained by the lack of a substrate that unlocks defensibility properties. By constructing AI deployments that inherently support legibility, plausibility, sourcing, and contestability, the framework aims to meet the demands of regulatory regimes like the EU AI Act and sector-specific standards (e.g., DIN SPEC 91484).

The paper explicitly acknowledges limitations: the framework does not solve ontology governance (authority for revisions must come from external institutional bodies) and does not extend to inherently interpretable models without modification. Future work is identified as the operationalization of the framework in PDA case studies to evaluate its impact on auditors' ability to document, contest, and revise AI-supported assessments.

1. Lifting: Translating "Pixel Speak" into "Human Speak"

2. Constraining: The "Reality Check" Filter

3. Typing: The "Trust Score" Tag

4. Revising: The "Undo Button" with a Paper Trail

The Big Picture

More like this