What is it about?
When you browse property listings online, some homes are marked as "brand new." This label matters it shapes what buyers see, how housing data is reported, and how cities track new construction. But these labels are not always accurate, and checking them one by one by hand takes too much time and effort. We built a system that does this checking automatically. It looks at the words used in a listing, details like the property size and completion date, and simple image clues for example, whether the listing includes architectural renders or floor plans, which are common for new developments. Because we rarely have enough verified examples to train a traditional AI model, our system learns from many small, imperfect signals instead of relying on a large set of hand-checked labels. One thing that makes our approach different is that it looks at groups of properties, not just individual ones. If several listings belong to the same building project, the system makes sure their labels make sense together rather than treating each one separately. The results are shown in an interactive tool where analysts can explore a map, see how listings connect to developers and agencies, and understand why the system flagged a particular listing. The goal is not to replace human judgement, but to help people focus their attention where it is needed most.
Featured Image
Photo by Artful Homes on Unsplash
Why is it important?
Spotting misleading or inaccurate property labels is a real problem for housing platforms, analysts, and policymakers but until now, most automated approaches have required large amounts of carefully checked data, which is expensive and slow to collect. What makes this work different is that it shows how to build a reliable system even when clean labels are scarce. Instead of waiting for enough verified examples, the system learns from many small, imperfect signals that already exist on any large platform text patterns, metadata rules, legacy flags, and image clues. This makes it practical to deploy in the real world, not just in research settings. The timing also matters. As housing markets tighten and demand for accurate supply data grows, the ability to automatically verify property claims at scale becomes increasingly valuable for regulators, researchers, and platform operators alike. Beyond real estate, the approach applies to any domain where structured claims need to be checked against noisy, mixed evidence such as product descriptions in e-commerce, content labels on media platforms, or compliance checks in healthcare. The interactive tool we built also addresses a common criticism of AI systems: rather than hiding decisions inside a black box, it gives analysts the context they need to understand, question, and act on the results.
Perspectives
This project started from a very practical frustration. Working with real property data, I kept seeing how a single mislabelled listing could quietly distort housing supply figures, mislead buyers, and slip through unnoticed because no one had the time to check every record manually. That felt like a problem worth solving properly. What I found most rewarding was building something that works with the messiness of real data rather than assuming it away. In research, we often start with clean, well-labelled datasets. In practice, those rarely exist. Teaching a system to learn from noisy, incomplete, and sometimes contradictory signals and still produce decisions that analysts can trust and interrogate felt like a more honest way to do applied AI. I also cared a lot about making the results explainable. It was never enough for the model to be accurate; it needed to show its reasoning so that a person could push back, override it, or decide to dig deeper. That balance between automation and human judgement is something I want to keep exploring. I hope this work is useful not just to people working in property, but to anyone dealing with imperfect labels and relational data which, in my experience, is most people working on real-world problems.
Kiana Bagheri Lotfabad
Macquarie University
Read the Original
This page is a summary of: Weakly Supervised Multimodal Claim Verification with Entity-Graph Inference, May 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3774905.3793141.
You can read the full text:
Resources
Contributors
The following have contributed to this page
