Detecting hosts who move bookings off the platform

The platform earns commission on bookings made through its website. Some hosts try to convince guests to book directly with them. With hundreds of thousands of messages in different languages, finding these cases manually is nearly impossible.

The goal was to identify these messages and the hosts involved.

The dataset contained host messages in multiple languages. Some messages included phone numbers or email addresses written in unusual ways to avoid detection.

Only rounded and aggregated results are shared here. No private messages or client data are included.

The pipeline uses three layers to detect suspicious messages:

• Rule-based detection: Finds phrases that suggest booking outside the platform.
• Contact detection: Identifies phone numbers, email addresses and hidden contact details.
• Machine learning: Uses multilingual sentence transformers to recognize suspicious messages even when they contain no obvious keywords.

Multilingual sentence transformers help the model understand messages across different languages.

Each message receives a score. These scores are then combined at host level to identify hosts who may be moving bookings outside the platform.

The pipeline was built using Python scripts and a Jupyter notebook. This makes it easy to run again with new data and improve the detection rules.

Analyzed over 320,000 host messages and identified an estimated €420,000 in commission at risk for the platform to investigate.