Beyond the Graph: Rethinking Viral Marketing for Physical Locations in LBSNs

Executive Summary

TL;DR: In the world of Location-Based Social Networks (LBSNs), a friend's recommendation is only effective if the target location fits into your physical orbit. This paper identifies the "Location Promotion Problem" and replaces static social influence models with a dynamic Location-aware Independent Cascade Model (LICM). By shifting the focus from "who you know" to "where you go" via Distance-based Mobility Models (DMMs), the researchers provide a robust framework for businesses to select the best "seed" users to spark a visiting trend.

Positioning: This work is a pivotal transition from purely topological influence studies (like the classic Kempe et al. ICM) to spatio-temporal behavior modeling, bridging social network analysis with human mobility science.

The "Static" Fallacy in LBSNs

Why does a viral tweet about a new restaurant in New York fail to bring in customers from Los Angeles? Traditional influence maximization assumes propagation probability ($pp(u,v)$) is a fixed value based on friendship strength.

In LBSNs, the authors argue that $pp(u,v|\ell )$ is query-dependent. It is not just about user $u$ influencing user $v$; it is about whether user $v$ is physically likely or willing to travel to location $\ell$. Existing models failed because:

1. They ignored Individual Mobility: Everyone has a unique "activity space."
2. They lacked Geographic Context: The distance between a user's current location and the target store is the ultimate bottleneck.

Methodology: The Power of Distance

The core innovation lies in the Distance-based Mobility Model (DMM). While prior works used Gaussian-based models (GMMs), these were often too rigid to capture the "long-tail" nature of human movement.

The Two-Layer Architecture of DMM

The probability of a user $v$ checking in at target $\ell$ is modeled as: $p(\ell |v)={\sum }_l{p}_{v,l}\cdot f_v(d(l,\ell ))$

1. Stationary Distribution ($p_{v,l}$): Represents the probability of user $v$ being at an existing visited location $l$. The authors use Random Walk with Restart (RWR) to establish these "home bases."
2. Distance Preference ($f_v$): A Pareto distribution representing how likely the user is to travel a certain distance. This captures the "Heavy-Tail" phenomenon: most movements are short, but occasional long-distance trips occur.

Category-Aware Sophistication

The authors further refine this by adding Categorical Preferences. If the target is a "Restaurant," the model weights the user's history and distance preferences specifically within that category, solving the data sparsity issue where a user might have many check-ins but few for a specific venue type.

Figure: Comparison of how DMM captures check-in behavior more naturally than Gaussian distributions.

Experiments: Real-World Evidence

The team tested their models on two massive LBSN datasets: Gowalla and Brightkite.

Key Findings:

• DMM vs. GMM: DMMs showed higher log-likelihood, meaning they predict real-world check-ins with significantly more accuracy.
• Influence Propagation: When tasked with predicting who would actually visit locations like Central Park (NYC) or Caltrain Station (SF), DMMs outperformed traditional static methods (like In-degree or Jaccard similarity) by a wide margin.
• Social vs. Locality: Interestingly, the study confirms that social influence in LBSNs is a "trigger" rather than a "driver"—a user's inherent mobility habits are the primary predictor of where they go next.

Figure: ROC Curves showing that Distance-based models significantly outperform static network benchmarks.

Critical Insight & Conclusion

The standout takeaway is the Physical Constraint of Influence. In the digital world (Twitter/Facebook), the cost of "accepting" an influence is a click. In the physical world (LBSNs), the cost is distance and time.

Future Implications:

• Precision Marketing: Businesses shouldn't just target "influencers" with high follower counts; they should target "local hubs"—users who frequently visit locations near the target venue.
• Limitations: The model assumes a stationary mobility pattern. It may not account for sudden shifts in behavior (e.g., travel/vacations) without a longer historical window.

By grounding social influence in the reality of physical distance, this paper provides a more scientific, quantifiable approach to location-based viral marketing.