Enhancing millions of rental listings with visual intelligence to improve discovery and quality
Problem
Flexibility is a key trend emerging from the pandemic as consumers become less tethered and the lines blur between travel, living, and working. Classifying listings by these visual attributes was hard given the sheer volume of unstructured data to turn into signal.
Solution
Labelbox Annotate and Catalog, with active learning and a shared ontology, let the company rapidly produce visual signal across millions of live property listings.
Result
The company rapidly built the ability to visually enrich rental properties into many unique listings, with a focus on venues outside popular tourist destinations. The application surfaces and drives demand to these listings, which have collectively earned more than $300 million globally since the start of the pandemic.
A leading vacation rental company needed to classify and enrich millions of listings from images. Labelbox automated the pipeline so models generate most of the signal, with experts accepting it, across nine million tasks.
The challenge
A leading vacation rental company wanted to standardize and automate its machine learning infrastructure so the whole organization could run repeatable data science. Business units across Operations, Customer Support, and Pricing handled unstructured data in different formats, and the priority was classifying, labeling, and enriching listing images — at the scale of millions of live listings. Flexibility was the emerging consumer trend, with the lines between travel, living, and working blurring, so classifying listings by visual attributes mattered, and the volume of unstructured data made it hard.
The approach
The company used Labelbox to produce the signal and automate the pipeline. Active learning prioritized the right data to label, tightly integrated with the labeling operation. A shared ontology gave every team a standard way to classify and label, and Issues and Comments made the work collaborative — so the company could create unique listings with rich metadata far faster. Labelbox Annotate and Catalog tagged and categorized millions of live property listings.
The outcome
After three months, the pipeline was fully automated: models generate most of the labels, and subject matter experts accept them. Human labeling costs dropped to a fraction of where they started, with over nine million annotation tasks completed in Labelbox. The application surfaces and drives demand to unique listings — especially venues outside popular tourist destinations — which have collectively earned more than $300 million globally since the start of the pandemic.
Where this goes
A model that generates its own signal, with experts in the loop to accept it, is a self-improving data engine. That's how you enrich millions of listings without millions in labeling cost.