Abstract
Predicting user activities and interests has many applications, for example, in contextual recommendations. Although the problem of predicting interests in general has been studied extensively, the problem of predicting when the users are likely to act on those interests has received considerably less attention. Such predictions of timing are extremely important when the application itself is time-sensitive (e.g., travel recommendations are irrelevant too far in advance and after reservations have already been made). Particularly important is the ability to predict likely future activities long in advance (as opposed to short-term prediction of imminent activities). In this paper we describe a comprehensive study that addresses this problem of making long-term time-dependent predictions of user interest. We have conducted this study on a large collection of visits to various venues of interest performed by users of Foursquare. We have built models that, given a user's history, can predict whether or not the user will visit a venue of a particular type on a given day. These models provide useful prediction accuracy of up to 75% for up to several weeks into the future. Our study explores and compares various feature sets and prediction methods. Of particular interest is the fact that venues interact with each other: to predict visits to one type of venue, it helps to use the history of visits to all venue types.
Introduction: The Challenge of Timing
Most recommendation engines excel at identifying user tastes but remain "time-blind." Suggesting a restaurant five minutes before a meal or a flight discount after a booking is ineffective. The core challenge addressed here is long-term, time-dependent prediction. Unlike "next-step" predictions that look only an hour ahead, this research focuses on predicting activity days or even weeks in advance—a critical capability for travel and event planning where decisions are made well in advance.
Methodology and Data
The researchers framed the problem as a binary classification task: will a specific user visit a target venue type on a given future date? To answer this, they analyzed a dataset of 180 million public Foursquare check-ins broadcast via Twitter, focusing on a subset of 55,000 "power users."
The study utilized Foursquare’s three-tier hierarchy to categorize venues:
- Tier 1: 8 broad categories (e.g., Food, Travel).
- Tier 2: 239 sub-categories (e.g., Italian Restaurant).
- Tier 3: 348 specific types (e.g., Pizza Place).
A key technical innovation was the use of "global" history. Rather than just looking at the target venue, the models incorporated visits to all venue types to capture hidden correlations between different activities.
Results: Interconnected Activities and Accuracy
The results demonstrate that human activities are deeply interconnected. Predicting a visit to a restaurant is significantly more accurate when the model considers the user's history with other venues, such as gyms or offices. Activities do not happen in a vacuum; they are part of a larger daily or weekly routine.
As shown in the table above, the models achieved an accuracy of 75% (at the Equal Error Rate point) for Tier-1 venues. Notably, adding features from "other" venue types improved performance by 3 percentage points (approximately a 10% relative improvement) over models restricted to the target venue's history.
Long-term Viability and Seasonality
Human behavior exhibits strong weekly seasonality. The researchers found that the error rate for predictions is influenced by these cycles, with models performing reliably even when looking several weeks into the future. The "N-Days Ago" approach allowed the model to rely on historical patterns from the same day of the week in previous months to make these long-term forecasts.
Even as the prediction horizon extends to 30 days, the models maintain a useful level of accuracy, significantly outperforming baselines that rely solely on simple averages. This long-term viability is essential for practical applications where decisions are made well in advance and are difficult to reverse once committed.
Conclusion
This study highlights that social network data contains robust signals for predicting future behavior. By accounting for the temporal context and the interactions between different types of venues, it is possible to move beyond "next-step" recommendations toward truly proactive, time-sensitive user assistance. Future work will explore latent activity patterns and the influence of social ties to further refine these predictions.
Summary updated