On the surface this seems like a very simple interactive recommender.
You are asked eight questions, that try establish whether you are formal or casual, interactive or an observer, about art or the wildernes, social or quiet…but all in a way that is relative to the types of things one might do on a date or a trip somewhere.
In the end you are given a suggestion for a city that you should visit, in my case that was Niarobi, Kenya.
This system is very interesting. AirBnB does not have some long history with me. They do not have a ton of demographics or a purchase history, so they ask a series of questions.
With five answers per question and eight questions, there are clearly a pretty wide-array of possible combinations, which could inform the engine.
AirBnB has a whole host of hosts and vacation packages, with descriptions and a lot of images and copy about those places.
My best guess on how this works is that it is a hybrid system.
With no user information in their database about past trips, ratings our user interests, AirBnb could seed the system with tags reflective of the type of adventure a given trip has to offer and this could be served based on the ‘item’ you describe in your personal wants for your free time in the questions. This would be a content-baased system (eventhough it is aimed at a user). The fictional item you create is compared to a real (and available and seasonal) item in AirBnB’s ecosystem.
With the data of many users already in the system, in the form of trip reviews - both stared and written, AirBnB has enough raw material (particularly if reviewers have answered the same questions you have) to provide a collaborative system where the location and the user are paired based on the experiences/ reviews of other users like you. Using nlp and stars, is a pretty good start with the items themselves, combined with results of your ‘quiz’ they can make suggestions.
If you have used their service before, are active in giving reviews and are active in the AirBnB community, your provided content could go a very long way in providing useful user information for a collaborative user-item system, but there may also be some diffusioin methods at work coming from your own written reviews.
It seems pretty clear that this is not a one-trick recommender system, but a complex interweaving of techniques. A hybrid of content and collaboration as your involvement becomes more they reliance on the quiz is less necessary.
Honestly this is quite brilliant of a solution for the cold start problem.
As for the success…I would say it is on the right track. In about four of the eight questions, I was torn between answers.
Would I prefer a 5-Star meal for a first date or a hike and a smoothie (hold the banana, yuk!) or a museum tour?
I think depending on the day I could pick any of those three. So, I did actually do so. I ran the quiz a few times, and depending on which of my interests I was focused on in question.
I was presented with some pretty relevant choices. Niarobi, San Francisco & Barcelona. Two of three are a hard yes! I would go without hesitation both Barcelona and San Francisco and they are in fact on my bucket list.
Niarobi, is not a likely destination, but the algorithm would have no way of knowing that that I am very cautious about traveling where there is high potential for crime, and there was nothing in the quiz which would reveal this. And honestly ifit were just a tiny be less risky, Nairobi would absolutely be somewhere I would go because of the geography, wildlife and uniqueness of the location.