Abstract
In recent years, Location-Based Social Networking (LBSN) sites such as Foursquare, Facebook Places, and Twitter have become extremely popular due to the extensive usage of location-enabled smartphone technologies. These LBSNs allow users to post check-ins that provide important information about activities and preferences. While existing research predicts user activities from social media data by considering time, venue, and occurrences, few studies investigate the influence of weather on these preferences. Psychological studies show that weather has a strong influence on human behavior. In this paper, we predict travel modes, day/night activities, and future visits based on weather conditions derived from social media check-ins. We develop several classification models to predict these behaviors using two real-world datasets from Tokyo and New York City. Our classifiers achieve substantial predictive strength, with an average AUC of 72.77% for the Tokyo dataset.
Introduction
While Location-Based Social Networks (LBSNs) have long been used to predict where a user might go next based on historical patterns, the environmental context—the "why"—often remains unexplored. One major missing piece of the puzzle is the weather. Psychological research consistently demonstrates that weather significantly impacts human decision-making, yet most predictive models ignore it.
This study addresses this gap by using a comprehensive list of weather parameters (such as humidity, wind speed, and precipitation intensity) to model and predict multi-faceted user behaviors. By fusing Foursquare datasets with weather API metrics and employing ensemble machine learning models, the researchers prove that our digital footprints are heavily influenced by the sky above us. Understanding these patterns is crucial for urban planning, targeted marketing, and improving recommendation engines.
Methodology
The researchers developed a multi-step pipeline to transform raw check-ins into predictive insights, as illustrated in the general methodology flow.
Data Fusion and Weather Integration
The core of the approach involves cross-linking Foursquare check-in data with the Dark Sky API (formerly Forecast.io). For every check-in, the system retrieves 13 specific weather attributes—including temperature, "feels like" temperature, wind bearing, and visibility—based on the exact latitude, longitude, and timestamp.
Feature Selection and Modeling
To identify which weather factors were most relevant, the team used Chi-Square ($\chi^2$) tests for categorical data (like weather icons) and Fisher’s Linear Discriminant Analysis (LDA) for numerical data.
Because real-world data is often imbalanced—for example, significantly more people check in on "Trains" than in "Private Transport"—the authors employed SMOTE (Synthetic Minority Over-sampling Technique). This generated synthetic instances of minority classes, preventing the models from becoming biased toward the most common behaviors. While several algorithms were tested, including Naive Bayes and REPTree, Random Forest Ensembles provided the highest accuracy across the four target models: Transport Mode, Day-time Activity, Night-time Activity, and Future Visit Prediction.
Datasets
The study utilized two primary real-world datasets:
- Yang’s Foursquare Dataset: Contains 573,703 check-ins from Tokyo and 227,428 from New York City.
- Dark Sky API: Used to augment the LBSN data with historical weather conditions.
By comparing Tokyo and NYC, the authors observed how different urban infrastructures (like Tokyo's highly efficient rail system) change how weather influences behavior.
Key Insights and Ideas
- The "Weather Icon" Effect: Interestingly, the high-level "weather icon" (a summary like "Rainy" or "Sunny") often had a higher correlation with user behavior than raw numerical data. This suggests that humans often make heuristic decisions based on a quick glance at a weather app rather than an analytical calculation of exact wind speeds.
- Infrastructure Matters: Models performed better on the Tokyo dataset than the NYC dataset. The authors hypothesize that because Tokyo’s public transit is more robust and predictable, user responses to weather follow more consistent, learnable patterns.
- The Humidity Factor: In Tokyo, humidity was the most influential factor for night-time activities. High humidity prevents sweat evaporation, leading to discomfort that drives users toward specific indoor nightlife spots or keeps them at home.
- Predicting the "Unpredictable": Future visits to specific venues like "Sea Beach" were highly predictable (AUC of 98%) because that behavior is almost entirely weather-dependent.
Key Results
The ensemble models demonstrated substantial predictive power:
- Transport Mode: Achieved an AUC of 85.1% for Tokyo after addressing class imbalances.
- Activity Prediction: Day-time activity prediction reached an AUC of 65.1%, while night-time activity reached 72.9%.
- Regional Comparison: Tokyo models generally outperformed NYC models (e.g., 75.25% vs 65.75% for future visit prediction), highlighting the role of local culture and city layout.
- Overall Performance: The average AUC of 72.77% represents a significant leap over earlier regression-based studies that often hovered around 25% accuracy.
Limitations and Future Work
Despite the success of the models, the authors acknowledge a few hurdles. People rarely "check in" when staying at home, making that specific class difficult to predict using LBSN data alone. Furthermore, the behavioral differences between Tokyo and NYC suggest that a "one size fits all" weather model is not feasible; models must be tuned to specific geographic and cultural contexts. Future work involves integrating these classifiers into real-time recommendation systems that can suggest activities or transport modes to users based on the current forecast.
Summary updated