Foursquare is a search and discovery tool which helps users discover venues around the world. Much of the data for these recommendations come from its sister app Swarm, which is a location based social network where users can \check in" to places they visit. Older versions of Foursquare had a strongly static component to its recommendations. For instance, the top restaurants in New York City do not vary from month to month, and venues with years of consistently strong signals will dominate search results. In this paper we outline a new algorithm which Foursquare uses in order to discover fresh recommendations. Promoting younger venues with fewer check-ins or older venues with a recent surge of activity increases turnover in our recommendations and yields a better user experience.
Foursquare has a database of nine billion check-ins and 85
million public venues around the world. Using this data, the
mobile app provides personalized venue recommendations to
users. Core components of these recommendations are based
on foot tra c data in the form of check-ins and passively
generated visits from a background location service called
Pilgrim [
There is a constant tension between consistency and freshness in Foursquare's recommendations. For example, Thomas Keller's Per Se is always at the top of the results for restaurants in New York City, but most users nd value in discovering a more accessible venue like a new mom-and-pop co ee shop around the corner. Likewise, a celebrity chef moving to a new restaurant results in a urry of activity which is not always captured well by Foursquare's long-term signals. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the authors.
RecSys ’16 Sept 15–19, 2016, Boston, MA, USA c 2016 Copyright held by the authors.
Foursquare has successfully implemented short term trend
detection to showcase real-time events as they happen [
All of the features described below are generated by users' interaction with the Foursquare and Swarm apps and by passively generated visits from Pilgrim. Noteworthy venues inspire users to interact with their apps, and so most user activity for a venue is seen as positive. 2.1
Checkins and visits: The primary signals for trendiness are based on foot tra c in the form of active check-ins and passive visits. Active check-ins typically indicate better venues, since Swarm users tend to broadcast special outings more often than their day to day activities.
Saves: Foursquare users have the option of saving a venue to a list for later. This distinguishes trendy new places from average ones, because it indicates aspirations to visit.
Tips from users: Users have the option of writing tips at
any venue, which are shown to other users as part of the local
discovery experience [
Tips from vetted accounts: A handful of user accounts are unusually in uential. For example, some celebrities and local blogs about food maintain active Foursquare accounts with tens of thousands of followers. Tips from these accounts drive foot tra c and are a leading signal of venue trendiness.
Explicit feedback: The Foursquare app prompts it users to leave explicit ratings|like, dislike, or neutral|about the places they visit.
Photos: The excitement of visiting a noteworthy venue is often re ected by our users documenting their visit with photographs. 2.2
For each of the activities listed above, we calculate two statistics.
The rst statistic is derived from tting a trend line through the time series of the activity. The signal that we use is given by the equation ^
Day where ^ denotes the slope of the trend line through the time series with 56 days of data, and ^ denotes the standard error of the estimate ^.
In Figure 1 we display the number of check-ins per day for two restaurants in Istanbul, Turkey. The trend lines for both time series (omitted from the gure) have similar slopes. Although the value of ^ is positive and similar for both venues, the value of ^ is lower for Suppa than it is for Ka e Pi. Hence the signal S for Suppa is larger than the corresponding signal for Ka e Pi. In general, venues with erratic or spiky activity do not bene t from one-time events for this class of signals.
The second statistic is a decayed sum of the activity, calculated with a half life of 56 days.
D = X cde d;
d where d is the number of days prior to the current day, cd is the total amount of user activity on that day, and = ln 2=56. Note that short half lives are associated with noisier data, and long half lives lead to a lack of freshness. For example, the venue Kafe Pi in Figure 1 has a spike in activity on Day 39, which would have dominated the signal if the half life were too short. Longer half lives have more stability, and we found that very long half lives lead to a lack of freshness in our recommendations. In our research, 56 days is the best balance for both stability and freshness.
The distribution of the S-scores is roughly bell-shaped, while the distribution of the D-scores has a long tail. In order to combine the two classes of scores, we normalize each signal to a Gaussian distribution using the function N = 1(r); where is the cdf of the standard N (0; 1) distribution and r is the relative rank, between 0 and 1, of the venue when compared to all other venues and sorted by a given score. We then combine the signals linearly with hand-tuned coe cients. The largest coe cients are associated with the S-score of tips left by vetted accounts|a sparse but strong signal |and the S-score of Pilgrim-generated visits. These two scores account for more than 60% of the nal signal.
The combined signal is now being used as a primary
component of venue recommendations, and is showcased in the
main Foursquare app and on the website in the form of
weekly billboard-style \Trending This Week" lists in major
metropolitan areas (Figure 2). It is also frequently covered
in articles which feature best-of lists for many cities [