[Strategic Network Theory] Mapping the "Hidden Homophily": How Topic Modeling and Geography Shape Our Social Links

TL;DR

Why do we follow certain people on location-based apps? This paper moves beyond simple "link prediction" to build a strategic structural model of network formation. By applying Latent Dirichlet Allocation (LDA) to unstructured biographies and tweets, researchers found that "hidden" similarities in interests—coupled with physical proximity—drive nearly 20% of social connections.

Background: The Shift from Static to Strategic

In traditional social network analysis, graphs are often treated as "given" (exogenous). However, real-world networks are the result of strategic individual choices. If I follow you, it's because I expect a utility gain (information, social status, or shared interests).

This paper positions itself at the intersection of Game Theory and Machine Learning, moving from "What does the graph look like?" to "Why did these two specific people decide to connect?"

The Problem: The Unstructured Data Gap

Existing models usually focus on "revealed homophily" (e.g., we both live in NYC). But "hidden homophily"—the fact that we both enjoy "open-source software" or "pop culture"—is often trapped in unstructured text like bios or tweets.

1. Prior work ignored the micro-foundations of utility maximization.
2. Computational limits made it hard to process billions of possible user pairs.

Methodology: The Structural Framework

The researchers proposed a model where a link $g_{ij}$ forms if and only if both users $i$ and $j$ perceive a positive marginal utility.

1. The Utility Function

The utility is defined as: $u_{ij}={\alpha }_0+{\alpha }_1{X}_i+{\alpha }_2{S}_{ij}+{\epsilon }_{ij}$ Where:

• $X_i$: Individual characteristics.
• $S_{ij}$: Pairwise similarity (The core of the study).
• ${\epsilon }_{ij}$: Random taste heterogeneity (Type I Extreme Value distribution).

2. Extracting Similarity via LDA

To turn text into math, the authors used Latent Dirichlet Allocation (LDA) to transform biographies into topic vectors. The similarity between two users ($bio\_sim$) is then calculated as the cosine similarity between these vectors.

Figure 1: Examples of biography and tweet topics extracted via LDA.

3. Four Dimensions of Proximity

• Biography (Topic Modeling): Matching interests from user profiles.
• Geography (Hometown distance): Measuring physical friction.
• Mobility (Co-check-ins): Shared tastes reflected in physical movements.
• Tweets (Social content): Real-time alignment of topics.

Empirical Results: Does Similarity Actually Matter?

Using data from Gowalla (35 million check-ins), the team estimated the model using Maximum Likelihood Estimation (MLE).

Table 2: Significant positive coefficients for co_checkin and bio_sim across different city and national samples.

Key Findings:

• Bio Similarity: Highly significant ($p<0.01$). Users with similar interests are much more likely to link.
• Physical Distance: Has a "decay" effect—as distance increases, the probability of linking drops.
• The 20% Rule: In their counterfactual analysis, the authors removed the "homophily" effect. The result? The number of formed links dropped by 20%, quantifying exactly how much our "likeness" drives our social structure.

Critical Insight & Future Outlook

This work provides a bridge between social science intuition and big data engineering.

Takeaway for Practitioners: If you are building a recommendation engine (for dating, professional networking, or social media), simply looking at "mutual friends" isn't enough. You must extract latent topics from user-generated content to capture the "hidden homophily" that drives long-term engagement.

Limitations: The model assumes pairwise independence (my link to you doesn't depend on your link to others), which ignores "triadic closure" (friends of friends). Future models will likely need to integrate Temporal Graph Networks to account for these higher-order dependencies.

Conclusion

By treating network formation as a stable matching process rather than a random event, Lee et al. provide a robust framework for understanding the digital social fabric. It's not just about where you are (GPS); it's about what you talk about (LDA).