Location Prediction: Communities Speak Louder than Friends

Jun Pang, Yang Zhang

Abstract

Humans are social animals, they interact with different communities of friends to conduct different activities. The literature shows that human mobility is constrained by their social relations. In this paper, we investigate the social impact of a person's communities on his mobility, instead of all friends from his online social networks. This study can be particularly useful, as certain social behaviors are influenced by specific communities but not all friends. To achieve our goal, we first develop a measure to characterize a person's social diversity, which we term 'community entropy'. Through analysis of two real-life datasets, we demonstrate that a person's mobility is influenced only by a small fraction of his communities and the influence depends on the social contexts of the communities. We then exploit machine learning techniques to predict users' future movement based on their communities' information. Extensive experiments demonstrate the prediction's effectiveness.

Introduction

Predicting human mobility—determining where a person will be at a certain time—is a cornerstone for urban planning, location-based advertising, and personalized recommendation systems. While it is well-established that social ties influence movement, most existing research treats a user's entire friend list as a monolithic block. This paper identifies a critical gap: social influence is not uniform. A user might visit a specific venue because of their "foodie" friends but would never go there with their "office" community. Previous models failed to distinguish between these contexts, leading to a loss of predictive power. The challenge lies in partitioning a user’s social network into meaningful communities and determining which specific community drives a person to a specific location at a specific time.

Methodology and Data

The researchers proposed a multi-step pipeline to move from raw social data to precise location prediction. The process begins with Community Detection using the Infomap algorithm to partition a user’s ego-network (their immediate circle of friends) into non-overlapping communities based on information flow.

Ego-networks of two users partitioned into distinct social communities, showing how friends are grouped into sub-structures.

To validate this approach, the study utilized two massive real-world datasets:

  • Gowalla: A legacy Location-Based Social Network (LBSN) dataset containing over 6.4 million check-ins, focusing on active users in New York and San Francisco.
  • Twitter: A dataset of geo-tagged tweets from late 2014 to early 2015, providing a modern look at social mobility with significantly more social ties per user than Gowalla.

Quantifying Social Diversity

A central innovation of this research is the introduction of Community Entropy (based on Rényi entropy). This metric quantifies the diversity of a user's social life. A high entropy value suggests a user interacts with many distinct, well-defined groups.

Distribution of community entropy across the user base, illustrating the variance in social diversity among users.

Counter-intuitively, the study found that users with higher social diversity (high community entropy) are actually easier to predict. Because their social circles are more distinct, the "signal" from a specific community becomes much clearer when they visit a location associated with that group. For every location a user visits, the model identifies the "influential" community—the one with the most frequent movement areas nearby.

Key Insights and Results

The analysis revealed several striking patterns in human mobility:

  • The "Rule of Few": Despite having many friends, most users are influenced by only a tiny fraction of their communities—usually just one or two.
  • Contextual Shifting: Influence is highly volatile based on time and space. The community influencing a user's lunch choice is often entirely different from the one influencing their evening activities.
  • Performance: The community-based approach showed a clear edge over existing benchmarks. In the Twitter dataset, the model was approximately 20% more accurate than models using randomly sampled friends. Furthermore, it significantly outperformed the state-of-the-art Periodic & Social Mobility Model (PSMM), achieving an accuracy of 0.76 compared to PSMM’s 0.55 in the New York Gowalla dataset.

Conclusion

This research demonstrates that social communities are a far more robust predictor of movement than general friend lists. While the current model assumes non-overlapping communities and requires a baseline of active user data, it opens new doors for understanding social behavior. Future work may extend this to overlapping community detection and apply these insights to other social behaviors, such as information diffusion or interest adoption across different social silos.

Summary updated

Abstract

Humans are social animals, they interact with different communities of friends to conduct different activities. The literature shows that human mobility is constrained by their social relations. In this paper, we investigate the social impact of a person's communities on his mobility, instead of all friends from his online social networks. This study can be particularly useful, as certain social behaviors are influenced by specific communities but not all friends. To achieve our goal, we first develop a measure to characterize a person's social diversity, which we term `community entropy'. Through analysis of two real-life datasets, we demonstrate that a person's mobility is influenced only by a small fraction of his communities and the influence depends on the social contexts of the communities. We then exploit machine learning techniques to predict users' future movement based on their communities' information. Extensive experiments demonstrate the prediction's effectiveness.
BibTeX
@inproceedings{pang2015location,
  author = {Jun Pang and Yang Zhang},
  title = {{Location Prediction: Communities Speak Louder than Friends}},
  booktitle = {COSN},
  year = {2015},
  doi = {10.1145/2817946.2817954},
}
Figure