TL;DR
Most location prediction models treat all your friends as equal influences, but this paper argues that our movement is actually dictated by specific social communities (e.g., coworkers vs. family). By identifying these influential sub-groups and measuring a user's "community entropy," the researchers developed a predictor that significantly outperforms traditional models by focusing on the social context of our mobility.
What Problem Does This Paper Address?
Predicting human mobility—where a person will be at a certain time—is a holy grail for urban planning, location-based advertising, and recommendation systems. While we know that social ties influence where we go, most existing research hits a wall by treating a user's entire friend list as a monolithic block.
The authors identify a major gap: social influence is not uniform. You might go to a specific bistro because of your "foodie" friends, but you’d never go there with your "office" community. Previous models failed to distinguish between these contexts. The challenge addressed here is how to partition a user’s social network into meaningful communities and determine which specific community is "driving" a person to a specific location at a specific time.
Methods
The researchers proposed a multi-step pipeline to move from raw social data to precise location prediction:
- Community Detection: Using the Infomap algorithm, they partitioned a user’s ego-network (their immediate circle of friends) into non-overlapping communities based on the flow of information within the network.
- Quantifying Social Diversity: They introduced Community Entropy (based on Rényi entropy). This metric quantifies how "diverse" a user's social life is. A high entropy suggests a user interacts with many different, distinct groups.
- Identifying "Influential Communities": For every location a user visits, the model identifies which community has the most "frequent movement areas" nearby. This community is labeled as the "influential" one for that specific spot.
- The Predictor: They built a logistic regression classifier using features derived from these communities, such as the distance to the community's frequent areas, community size, and "community connectivity" (how tightly knit the group is).
Datasets
The study utilized two massive real-world datasets to validate their theories:
- Gowalla: A legacy Location-Based Social Network (LBSN) dataset containing over 6.4 million check-ins. The authors focused on "active users" (those with >100 check-ins) in New York and San Francisco.
- Twitter: A self-collected dataset of geo-tagged tweets from December 2014 to April 2015. This provided a more modern look at social mobility, featuring significantly more social ties per user than the Gowalla set.
Key Insights and Ideas
- The "Rule of Few": One of the most striking findings is that 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 that influences your "Wednesday Lunch" (likely colleagues) is almost entirely different from the one influencing your "Wednesday Dinner" (likely family or close friends).
- Diversity Breeds Predictability: Counter-intuitively, 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.
- Communities > Friends: The authors proved that "virtual communities" (randomly sampled groups of friends) perform much worse in prediction tasks than the naturally occurring communities detected by their algorithms.
Key Results
The community-based approach showed a clear edge over existing benchmarks:
- Superior Accuracy: In the Twitter dataset, the community-based model was roughly 20% more accurate than models that used randomly sampled friends.
- Outperforming PSMM: The model significantly beat the state-of-the-art Periodic & Social Mobility Model (PSMM). For instance, in the New York Gowalla dataset, the community model achieved an accuracy of 0.76 compared to PSMM’s 0.55.
- Strong Correlation: There was a very high Pearson correlation (up to 0.97 in some sets) between a user's community entropy and the model's prediction success (AUC), proving that social structure is a robust predictor of movement.
Limitations and Future Work
While the results are promising, the authors acknowledge a few areas for growth:
- Non-Overlapping Assumption: The current model assumes a friend belongs to only one community. In reality, social circles often overlap (e.g., a coworker who is also a college friend), and extending the model to "overlapping community detection" could improve accuracy.
- Data Sparsity: Like all mobility models, this requires a baseline of "active users." Predicting the movement of "lurkers" or infrequent check-in users remains a challenge.
- Broadening the Scope: The authors suggest that this community-centric view could be applied to other social behaviors beyond just location, such as how information spreads or how interests are adopted across different social silos.
Summary updated