The late journalist David Carr once asked at SXSW, "What the fuck is an algorithm? What does one feed it? How do you take care of it?" At its most simple, an algorithm is a step-by-step process that solves a problem, and online, algorithms rule everything. On Facebook, an algorithm is responsible for sifting through the 15,000 potential news stories from family, friends, and people you follow, narrowing them down to a manageable and hopefully meaningful number of stories that show up in your News Feed. It tells Amazon that after I've searched for The Fast and the Furious, I might also like to watch the second film (Own it on blu-ray! Better yet, get the boxed set!). For those who stream their entertainment online, algorithms help analyze the data of millions of viewers and their habits, and even dictates what shows Netflix produces, like House of Cards and Fuller House.
How effective are algorithms when it comes to recommending something as culturally ingrained and intensely personal like food?
Unsurprisingly, algorithms are also responsible for fueling restaurant apps. Restaurant recommendation (or "discovery") apps — from stalwarts like Foursquare to upstarts like MyFab5 — come up with recommendations based on algorithms' interpretations of data (yours and theirs). Others, like the just-launched Luka and the six-month-old Flavour, apply algorithms to recommendations curated by professionals and tastemakers. But when biases, context, and emotion are removed from the equation, how effective are the algorithms when it comes to recommending something you'd like — especially when it's regarding something culturally ingrained and intensely personal like food? And how do you find a balance between the two?
How do you improve an algorithm?
Algorithms are constantly tweaked and refined so businesses can better interpret the mountain of raw data used to help them sell to you and to sell you. But when collecting user data, the algorithm does not always consider context and human behavior — "your" data could be skewed if you're sharing an Amazon Prime/Netflix account with a friend or shopping for a gift. As a result, users can unwittingly "outwit" the recommendation algorithm.
The six-year-old discovery app Foursquare comes up with its recommendations through "collaborative filtering," which looks at users' previous browsing history, purchases, or check-in history to personalize results. According to a 2011 blog post by a Foursquare engineer, the app's early versions of the recommendation app built collaborative filtering on top of a "cold start" algorithm that pulls venues based on the new user's geographic location and "social encouragement" — where their friends and fellow app-users have already been. For new users, building the personalized recommendation algorithm starts with the first check-in. "We can go beyond 'X people like this spot' and look into the other types of places that people like that spot go," says Foursquare representative Laura Covington.
And after accumulating seven billion check-ins' worth of user data, Foursquare (which recently spun-off its "check-in" feature into a separate app called Swarm) has a pretty good idea of where to go and what to do, even if you're a new user. "If you went to Roberta's (a popular pizza place in Bushwick, Brooklyn), we can learn more than just you like pizza: We also may infer that you like artisanal pizza with local ingredients, and you might like cocktails too," Covington says. "So, we might recommend someplace like Paulie Gee's (a popular pizza place nearby in Greenpoint, Brooklyn) next time you are looking for pizza. And we might suggest that you head to the Narrows (bar in East Williamsburg, Brooklyn) for a cocktail, because there are strong connections between the people that go to these places."
Foursquare currently places a stronger emphasis on discovery and finding recommendations with better filtering. "We're getting closer and closer to contextual recommendations," Covington says. "Our first version of real-time recommendations, called Radar, was introduced in 2011. We quickly realized that hardware just wasn't advanced enough to do what we wanted and pinpoint your exact location so that we can send the right notification. So we spent a couple of years solving the problem through software." A new version of real-time recommendations rolled out in 2013 under the name Pilgrim: The upgrade kept track of users' locations even while the app was not in use, sending push-notification recommendations.
Last summer, after spinning off Swarm, Foursquare began asking users about things they liked, or "Tastes," in order to provide recommendations that "continue to get smarter and a little more human through small refinements," Covington says. If you love pimento cheese, it could recommend Brooklyn Kolache Company for a pimento cheese breakfast pastry. Foursquare's new focus is on curation from "experts" within the app. This is especially helpful in sparsely populated areas, where only a handful of users might dictate what check-ins are popular, or if other scenarios might skew the collected data. (For example, when driving cross-country, more Foursquare users might be checking in to a convenient fast-food restaurant off the highway instead of a smaller mom-and-pop spot.)
"We have a data team that is constantly trying to solve the problem of finding the newest and greatest places around the world," Covington says. "You can imagine the scale of this problem — instead of finding out about cool places with an ear to the ground, we are working on doing it with machine learning."
Are expert-filtered algorithms better?
Foursquare's "Trending this month" feature is particularly helpful, but the idea of a successful algorithm filtered through "recommended by experts" depends on the experts listed. Luka is touted as a cross between Siri and Yelp and recommends restaurants the same way you might ask your friends — via text. Currently only available in San Francisco (coming soon to New York), users text questions about restaurants within the app and Luka responds with a suggestion. The user can respond again either by coming up with a different question (like popular items, hours, WiFi, other restaurants nearby), or use one of the pre-programmed responses to continue or end the conversation.
The current lack of human oversight becomes an issue, not an asset.
According to Founder and CEO Zhenya Kuyda, Luka's algorithm prioritizes restaurants based on the number and quality of reviews (a professional review would count more than a typical user review). Better yet, she says, "Luka remembers what you don't like and what you like and takes it into account for future recommendations. When Luka learns that you went to a place — usually if you decide that you're going there while getting recommendations or when you say explicitly 'I've been there' — it starts asking questions about your experience and what you liked, if anything." The app is beautiful, simple, and presents options outside of a standard list view like Yelp or Foursquare. But it's not quite Her.
With Luka, you might get recommendations the same way you would by asking your friends: via 10 disparate text messages that still don't really tell you where you want to go. While Luka is certainly faster than a recommendation culled by sifting through online reviews, it needs some fine-tuning: Given the disparity between the number of restaurants and professional reviews, the current lack of human oversight in Luka becomes an issue, not an asset. Asking about Thai food in San Francisco's Union Square (with the popular Kin Khao in mind), Luka suggested an Italian coffee shop. Asking about a newer French restaurant did not give any new results, like the recently reviewed Monsieur Benjamin. In order to compensate for lack of reviews for newer restaurants, Luka does send users notifications about new opening restaurants, and even offers to make a reservation.
In "human recommendation" apps, diversity of opinions is key. YP-Dine, a Yellow Pages-backed food and drink app that recently launched in Montreal, provides recommendations based on feedback given by an internal team of editors and tastemakers. If you're not confident in the recommendations of someone you're not familiar with, Chefs Feed eschews this model by providing recommendations from well-known chefs. Upon entering the app, users are provided with slideshows of recommended dishes near their location, featured dishes, and featured chefs. Users can also follow chefs, receiving updates when they post recommendations, introduce new dishes, and post announcements. Users can add recommended dishes to a "to-do" list; once they're tried, dishes can be rated to improve the app's recommendations.
Most apps will only be as good as the user's input and needs.
Chefs Feed co-founder Jared Rivera pinpoints problems with both collaborative-filtering algorithms and "curated" recommendations. "When looking at Panda Express on Yelp, it's no surprise that Yelp's algorithm tells me that people also reviewed China Express," Rivera says. But the problem in recommendations, he says, is that even though the quality of dishes in his chef-recommended app "can be objectively reviewed by somebody who understands the mechanics of cooking... the taste of the dish itself is still subjective for every diner." That said, Rivera says Chefs Feed v4.0 (to be released later this year) will allow chefs to easily add their favorite dishes directly to the platform. Chefs Feed also plans to refine its algorithms to help users identify dishes they like based on dishes they've expressed affinity toward. It's a finer recommendation algorithm, distilled through content provided by chefs, and user feedback.
Everyone is looking to distill recommendations for every occasion, but most apps will only be as good as the user's input and needs. An app like Foursquare or Luka might be better when looking for places to check out in a specific neighborhood, while Chefs Feed and YP Dine are better suited for tourists or for trying a new or overlooked dish. The constantly evolving restaurant scene calls for a constantly changing algorithm, and a constantly updated platform, too.
According toThe New Yorker, Netflix's Ted Sarandos suggested during a Sundance panel that algorithms "[I]n practice, [are] probably a 70-30 mix. Seventy is the data, and 30 is judgment... but the 30 needs to be on top, if that makes sense." The data is worth holding on to, but it needs at least a little of that human touch.