Multiple Location Profiling for Users and Relationships from Social Network and Content

Rui Li, Shengjie Wang, Kevin Chen-Chuan Chang

Abstract

Inferring the geographic locations of social media users is critical for personalized services, yet most users do not share explicit GPS data. Existing methods often assume a single "home" location, failing to account for the reality that users often have multiple significant locations (e.g., home, work, college) and engage in "noisy" social behavior unrelated to their physical presence. This paper introduces the Multiple Location Profiling (MLP) model, a generative probabilistic framework that discovers a user's various geographic footprints by modeling social networks and tweet content while explicitly filtering out non-geographic noise. MLP significantly improves home location prediction and successfully identifies secondary locations and the geographic context of social relationships.

The Challenge of Mixed and Noisy Signals

Knowing a user's location is highly valuable for personalized search, local news delivery, and targeted advertising. However, very few social media users share explicit GPS coordinates. Prior research attempted to infer a user's location by looking at their friends' locations or the local words in their posts.

Existing methods suffer from two major flaws:

  1. The Mixed-Signal Challenge: Users often have multiple locations (e.g., a hometown and a college town), and their social networks and tweets reflect all of them.
  2. The Noisy-Signal Challenge: Users follow celebrities or tweet about places unrelated to their physical location (e.g., following Lady Gaga or tweeting about a vacation to Honolulu).
Figure 1: Conceptual overview of the MLP model. It identifies Carol's multiple locations (Los Angeles and Austin) by analyzing her network and tweets while filtering out noise like vacation plans or celebrity follows.

Forcing this complex data into a single-location bucket leads to inaccurate profiling. The Multiple Location Profiling (MLP) model addresses this by recognizing that a user's digital footprint is a mixture of multiple real-world locations and random noise.

The MLP Model

MLP is a generative probabilistic model conceptually similar to Latent Dirichlet Allocation (LDA). It assumes that observed data—tweets and follow relationships—are generated from hidden variables representing the user's true locations.

Key innovations include:

  • Location-Based Generation: The model assumes the probability of two users following each other decays with physical distance. Tweeting about specific venues is modeled as a distribution over locations.
  • Model Selectors for Noise Filtering: To handle noise, MLP introduces a "model selector" variable. For every tweet or follow, the model decides if the action was generated based on the user's physical location or if it was random noise.
  • Partial Supervision and Efficiency: MLP uses self-reported home locations from a small subset of users to anchor the model. To remain computationally feasible, it uses "candidacy vectors" to prune irrelevant cities from the search space during inference.

Key Insights: The Geography of Twitter

A critical finding of this research is that Twitter is less geographically bound than other social networks like Facebook. The authors found that the probability of a follow relationship on Twitter decays with distance following a power-law exponent of -0.55.

Figure 5: Following probabilities versus physical distance on Twitter, showing a power-law distribution fitting with an exponent of -0.55.

This exponent is significantly lower than the -1.0 observed in previous Facebook studies, meaning Twitter connections are much less sensitive to physical distance. This makes location prediction on Twitter inherently harder and necessitates robust models like MLP that can distinguish between local and non-local signals.

Key Results

The MLP model was evaluated on a large-scale Twitter dataset of over 139,000 users.

  • Home Location Prediction: MLP accurately predicted home locations for 62.3% of users within 100 miles, significantly outperforming network-only (52.4%) and text-only (49.7%) baselines.
  • Multiple Location Discovery: For users with multiple real-world locations, MLP achieved a distance-based precision of 50.6% and a recall of 47.0% for its top two predictions.
  • Relationship Explanation: MLP correctly identified the underlying geographic context of specific follow relationships with 57% accuracy, compared to just 40% for naive baselines.

Conclusion

By explicitly modeling noise and multiple locations, MLP provides a more nuanced and accurate geographic profile of social media users. While the model currently relies on external gazetteers and self-reported profile data, it opens the door to automatically grouping social networks into meaningful real-world clusters, such as separating college friends from professional colleagues based on their geographic context.

Summary updated

BibTeX
@article{li2012multiple,
  author = {Rui Li and Shengjie Wang and Kevin Chen-Chuan Chang},
  title = {{Multiple Location Profiling for Users and Relationships from Social Network and Content}},
  journal = {Proc. VLDB Endow.},
  year = {2012},
  doi = {10.14778/2350229.2350273},
}
Figure