Introduction
While location data is invaluable for applications ranging from local news recommendations to epidemic tracking and disaster response, it remains notoriously sparse on social media platforms. Fewer than 1% of tweets are natively geo-tagged, and user profile "location" fields are frequently empty, sarcastic, or intentionally vague. This research addresses the challenge of "city-level" granularity—a difficult task due to the vast search space—by introducing a system that infers a user's primary residence through an analysis of tweet content and temporal posting behavior.
Methodology: The Dynamically Weighted Ensemble
The researchers developed a sophisticated ensemble algorithm that moves beyond "one-size-fits-all" models. The system utilizes a Dynamically Weighted Ensemble of specialized classifiers:
- Content-based Statistical Classifiers: These employ Bayesian models to analyze general vocabulary, hashtags, and specific place names identified via the USGS gazetteer.
- Heuristic Classifiers: These track mentions of specific cities/states and scrape Foursquare check-in URLs to map user movements.
- Behavior-based Time Zone Classifier: This novel component analyzes when a user tweets, identifying temporal signatures (e.g., activity peaks after work hours) to predict time zones.
- Hierarchical Approach: The model first predicts a broad region (e.g., time zone) before selecting a city within that subset, significantly reducing noise.
A key innovation is the "Traveler Detection" binary classifier. By filtering out frequent travelers whose data is inherently noisier, the system improves accuracy for stationary users.
Experimental Setup and Datasets
The study utilized two primary datasets for evaluation:
- Primary Dataset: 1.52 million tweets from 9,551 users across the top 100 US cities, collected in mid-2011. Ground truth was established using initial geo-tagged tweets.
- Benchmark Dataset: A dataset from Cheng et al. [2010] containing over 5 million tweets was used for head-to-head comparison with existing state-of-the-art methods.
Results and Discussion
The hierarchical ensemble achieved 64% accuracy for exact city-level matches. Performance increased at broader granularities, reaching 66% for states and 78% for time zones. Notably, when identified "travelers" were removed from the dataset, city-level accuracy improved to 68%.
The research highlights that "temporal signatures"—the rhythmic patterns of when users wake and sleep—are powerful indicators of location that are harder for users to mask than text content. Furthermore, the "Dynamic Weighting" strategy allows the model to prioritize specific signals (like a city name) over common stop words.
Conclusion and Future Directions
The proposed system is highly efficient, requiring approximately 670ms to predict a location based on 200 tweets. However, limitations remain, including a current focus on US-centric data and the exclusion of social graph information (follower/following relationships). Future work will explore neighborhood-level prediction and the integration of social network signals to further refine accuracy.
Summary updated