Christina Cacioppo : Preferences revealed by data

Economist Paul Samuelson published his theory of revealed preference in the late 1930s. Revealed preference theory basically allows economists to calculate the preferences of a consumer – what that consumer considers his/her best available option in a given situation – using his or her purchasing behavior. Revealed preference theory has been influential to consumer choice theory because it allows economists to get a relative sense of preferences without mapping a consumer’s entire utility function.

This is a big deal because if you want to know what I’m going to like, or if I’ll like thing A more than thing B, you can just look at what I’ve chosen in the past rather than trying to figure out what I think of goods A and B at every price point.

I’ve begun to realize there’s a parallel with consumer web services and the “data exhaust” they generate: namely, that you can use my past behavior on some web services to predict what I will prefer on entirely different web services.

Let’s look at a two-company example of Foursquare and Yipit. Foursquare is a mobile service that allows me to check in to locations I’m at – businesses, parks, airports, offices, etc. – while Yipit is a daily deal aggregator. Like other daily deal sites, Yipit sends me one email each day with a deal they think I’ll find relevant, based on a few things I’ve told them about my preferences. (Screencap below.)

Selections like Yipit’s aren’t ideal: I have to fill them out, which takes time; there’s not the granularity I’d prefer among business types, times, locations (and if there were, there’s a good chance the interface would be complicated); and I have to tell the system explicitly what I want, so my answers will be subject to conscious and unconscious biases.

Suppose I’m interested in discounts on tanning services, but only in January and February, when the weather’s cold, and I feel pale. I could log on to Yipit, select “Tanning” in early January, peruse the deals that hit my inbox, log back on in March, and de-select the Tanning box – but c’mon. That all takes time, and it comes with cognitive load: I have to remember to do it. I have no access to data from Yipit, but I’d bet that preferences users set when they sign up are rarely changed. I believe first choices often become rules, which is why defaults and initial experiences are so important.

Enter Foursquare, which records places I’ve said I’ve been at specific times. If you’re a Foursquare user, log in to the website and check out this page: it’ll give you a preview of what the API spits back as your checkin history. Today, I can retrieve 283 days’ worth of my data, which amounts to 987 checkins at 471 unique locations. For any location, I can pull how often I visit (frequency), when I visit (times/days/months/seasons), and how many other people have checked in (popularity.)

I’d much prefer my Foursquare data power Yipit’s selection algorithm:

1. Foursquare gives more granular data than anything I’d have the patience to enter myself.

2. Foursquare knows where I go and where I don’t, which gives it the ability to target location within a few blocks. Courtesy of Where Do You Go, here’s a snapshot of my New York City checkins. Foursquare knows I spend 95% of my time in New York in the East Village or the Union Square/Flatiron area. I may want restaurant deals, but if Yipit gives me a restaurant deal in Tribeca (or worse, the Upper West Side) I’m never going to take it.

3. Yipit’s deal targeting becomes based on the types of venues I’ve been – and assumedly like – not what I say I like. Targeting drops down to “sushi places in the East Village” or “Italian sit-downs in Soho” rather than “New York City restaurants.”

Foursquare and Yipit represent a base or simple idea of what happens when data (or data exhaust) from one service is routed to another service. There are many, many more interesting things to do here.

At the end of this post – well, I want to say that I know and immensely respect the Yipit team. They’re really smart and dedicated, and they’re thinking of all of these things, plus more. I have no doubt they’re going to be successful.

Anonymous

I still have yet to be convinced that location based services will be taken on by more than just early adopters. LBS seems to do really well in densely populated areas with large selection opportunities, but until I see people who don’t live on (or near) a coast of this country and are older than 30 using these services I remain skeptical. Would love to hear or see some data pointing the other way.
http://www.christinacacioppo.com Christina Cacioppo

I agree – fair points with no clear answers for any of the location-based services, big or small.

I didn’t mean to reply as heavily on Foursquare + Yipit as “the most awesome thing ever” – although it may have come off that way. (Thanks for that feedback – helpful.)

More meant that I think there’s a big opportunity here in leveraging data across services to improve personalization and overall user experience.
http://www.jeffheuer.com Jeff Heuer

This is exactly the value proposition behind FOURCAST, the location-based recommendation platform I built on top of Foursquare and other checkin services. As you allude to, consumers are generating a lot of potential value via their data exhaust, but today that value is largely locked-up, and rarely returned to consumers in a meaningful way. FOURCAST turns your checkin history into a sort of abstract profile, which can then be used to generate personalized recommendations for places you haven’t been yet but are likely to enjoy. I think this type of “passive personalization” hits the sweet spot of delivering the most value to users while requiring the least amount of effort from them.
http://www.brekiri.com/blog/ Greg4

The data analysis approach can be much more powerful, like you mentioned, but it’s also dangerous to infer too much. On Foursquare, what if I would love a change of pace in a different neighborhood but just don’t usually make it? The system probably can’t tell. Or I might move or switch jobs. Or I might get tired of beer and decide that I’m interested in wine bars. Phenomena like “my Tivo thinks I’m gay” are great examples of the limitations of such analysis.

Ideally, this information would be automatically generated but manually editable, providing most of the best of both worlds. Of course, translating the data-driven version into a UI, and vice versa, is a little trickier.
http://viniciusvacanti.com Vinicius Vacanti

Great post Christina. Thanks for writing about Yipit and this is definitely something we are working on.
http://www.facebook.com/philip.j.cortes Philip Cortes

I completely agree! Duplicating efforts is more work for the user, and doesn’t translate to additional benefits. Should YipIt introduce machine learning into their algorithm, they could then create a powerful amount of value for their user base by leveraging Foursquare’s data, requiring no work from their user base.
Anonymous

I would add two points:

1. There’s a lot of cognitive load, as it is, to register those 987 foursquare check-ins. I often forget, my phone dies, don’t want to put out my phone in a social gathering, etc. I hope the time of smart implicit check-ins (e.g. I am there for 10 minutes and a friend on 4sq/FB is there too –> check me in) comes soon. Game mechanics only help so much. What’s great about other tools in the quantified self app (like FitBit, RunKeeper) is that they don’t need to be maintained. Foursquare does.

2. Christina says she’d never take a restaurant deal in TriBeCa. I am not convinced that past history is a self-sufficient predictor of future preferences and behavior. Christina is never in TriBeCa at dinner time, but is sometimes there after dinner (complete speculation from the heat map above, and in fact the time of day of the check-ins would greatly enhance the data), she might be interested in dinner there. Or maybe she does not want to have dinner in TriBeCa, but if the restaurant is exactly what she’s always wanted to try or anyhow piques her interest, she might want to go. The fact that is off the beaten track could flip and be a plus. Such a scenario would be improbable for the UWS, where the restaurant offer and the overall demographics are totally different, so that excluding it for the offers might make sense.

Apple and behavioral economics proved in the last ten years that, by far and large, we don’t know what we want, and we’re even worse when asked to explain why we want what we say we want. I think the taste graph than Hunch (and maybe others) are trying to build would be an essential integration to the 4sq data in order to present what might be relevant, and that hidden Markov models will be the grammar of data-driven marketing. No one knows what Christina wants, but we can make data-based inferences and create a portfolio of diversified offers for her.

Also, do I win something for the most words starting with a ‘w’ in a row?

Preferences revealed by data

Archives