Now consider another asset in the peer-to-peer economy — an Airbnb rental. What should you charge for one? The answer turns on a much longer list of variables beyond how many homes are available and who wants them. Is it a townhome, a penthouse, a cabin, a castle, a teepee, a yurt? A single room or a whole home? The bed: double or queen? The view: riverfront or city skyline? The Left Bank or right one? How gourmet is the kitchen? Is there a subway stop nearby or off-street parking? Are the Grateful Dead in town? Or the cherry blossoms blooming?
What's worth more: A studio in Adams Morgan with half a dozen bars nearby, or a single bedroom in a Capitol Hill Victorian with views of the Capitol dome?
The housing market and hotel industry wrestle with some similar questions in determining the value of a property. But Airbnb's puzzle sits at the incredibly complex intersection of the two, where every quasi-hotel room has the individual character of a single home, and every home has the seasonal fluctuations of the tourist industry.
Of course, one way to answer this question, if you rent your home on Airbnb, is to test a bunch of rates until you finally figure out what the market will bear. But Airbnb, a certain breed of big-data-loving tech company, wants to offer precise guidance to more than a million hosts. And that means cramming all of these questions into an algorithm.
Earlier this summer, Airbnb released the machine-learning platform, called Aerosolve, that the company has built to do this so that other developers might experiment with it. And this past week, Airbnb product lead Dan Hill published a lengthy and fascinating look for laymen at how the algorithm works in IEEE Spectrum, a publication of the Institute of Electrical and Electronics Engineers.
The platform incorporates image analysis from the photos hosts post (the warm colors in a bedroom influence how inviting it looks). It automatically divides cities into neighborhoods and micro-neighborhoods (Airbnb eventually realized that distance alone is a bad measure of similarity; nearby neighborhoods on opposite sides of London's Thames River command very different prices). The company also wanted to create dynamic pricing tips that vary by the day and mimic how hotel and airline prices rise and fall depending on when you book.
The system, Hill writes, considers hundreds of attributes, calculates probabilities at different price points, gives hosts tips and then later looks back to check if the tip worked:
Here’s where the learning comes in. With knowledge about the success of its tips, our system began adjusting the weights it gives to the different characteristics about a listing—the “signals” it is getting about a particular property. We started out with some assumptions, such as that geographic location is hugely important but that usually the presence of a hot tub is less so. We’ve retained certain attributes of a listing considered by our previous pricing system, but we’ve added new ones. Some of the new signals, like “number of lead days before booking day,” are related to our dynamic pricing capability. We added other signals simply because our analysis of historical data indicated that they matter.
The algorithm now reflects all kinds of quirky things that Airbnb has learned about its customers, things a reasonable person might not think would influence the value of a property. People, for instance, don't want homes in Seattle without WiFi connections (we can assume the imperative is different for remote cottages in the North Carolina woods). Listings that show photos of cozy bedrooms do better those with stylish living rooms. Guests are willing to pay more for a rental that has a lot of reviews.
A really, really expert human, in other words, probably couldn't get this precise.