That’s indeed a very interesting perspective. I believe some companies already start doing this — using computer vision to detect things that can be sold online.
We productionized the model using our internal tools. Some of them will likely be open sourced pretty soon. Here are some links about the current status of those tools: https://databricks.com/session/bighead-airbnbs-end-to-end-machine-learning-platform and https://databricks.com/session/zipline-airbnbs-machine-learning-data-management-platform. Hope this is helpful!
It has some legacy reasons in production. One of the benefits of building multiple binary models was that you didn’t need to recompute probability scores for existing room types. In retrospect this benefit doesn’t outweigh the problems it brings so we are going to productionize models in multi-class fashion.
Thanks for the comments! For Q1, the answer is no. But currently precision has much higher priority so we kind of settled with 50% recall. In the future, we definitely want to push recall much higher while still keeping high precision. For Q2, I’d say most of them are probably there. Your tool looks interesting and I will take a look.