[HT 2018] Beyond Words: Predicting Socioeconomic Status via Twitter Social Graph Embeddings

TL;DR

Can we predict how much you earn or what your job is just by looking at whom you follow on Twitter? This paper proves we can. By transforming the Twitter follower network into low-dimensional Graph Embeddings, the authors outperform previous text-only state-of-the-art models. The method is particularly revolutionary for "silent" users—those who read but never tweet—enabling accurate profiling without a single word of text.

Problem & Motivation: The "Silent User" Dilemma

In the realm of Computational Social Science, inferring latent attributes like income and occupation is a "Holy Grail" for personalized ads and political campaigning. Traditionally, researchers looked at what people say (Natural Language Processing).

However, the authors identify two critical flaws in this "Text-First" world:

1. The Inactivity Gap: Roughly 33% of Twitter users are inactive (posting fewer than twice). Text-based models are useless for them.
2. Missing Context: Social status isn't just about vocabulary; it's about the circles you navigate. The Homophily Principle suggests that "birds of a feather flock together"—people with similar incomes often follow the same influencers, news outlets, and peers.

Methodology: Turning Random Walks into Social DNA

The authors bridge the gap between Network Science and NLP by treating the social graph as a language.

1. Generating the "Social Sentences"

Instead of analyzing tweets, they perform Random Walks on the follower graph. Imagine starting at a user's profile and randomly jumping to one of their "follows." By repeating this 80 times, they create a sequence of User IDs. This sequence is functionally equivalent to a "sentence" in NLP.

2. The Skipgram Model

Using these sequences, they train a Skipgram with Negative Sampling (SGNS) model. The goal is to maximize the probability of predicting a user's "neighbors" in the graph.

The result is a 32-dimensional vector (embedding) for every user. In this vector space, two users who share similar follow patterns—even if they've never met—will be positioned close together.

Experiments & Results: Network vs. Text

The researchers tested their method on a dataset of 4,625 users mapped to the UK's Standard Occupation Classification (SOC) and mean yearly earnings.

Key Findings:

• Graph > Text: Graph embeddings alone (Graph) outperformed word clusters (Topics) in almost every model configuration.
• The Power of Combination: The "Graph+Topics" model achieved the best results, proving that who you follow and what you say are complementary signals.

Visualizing Social Homophily

The t-SNE plot below reveals clear clusters of high-income (black) and low-income (white) users. This visual evidence confirms that the social graph naturally "stratifies" by socioeconomic status.

Critical Analysis & Conclusion

Takeaway

This paper is a milestone because it shifts the focus from content to context. It demonstrates that for complex social traits, the "Interest Graph" is a more robust signal than the "Conversation."

Limitations & Future Work

• Compute Costs: Generating random walks for millions of users involves significant overhead compared to simple text feature extraction.
• Privacy Concerns: While the authors use public data, the ability to profile "inactive" users raises ethical questions about "invisible" surveillance.
• Platform Specificity: Twitter is an interest graph. Would this work as well on Facebook (a friendship graph) or LinkedIn (a professional graph)? The authors suggest that the homophilic bias might be even stronger in those environments.

Final Thought: Next time you hit "Follow," remember—you are not just curating your feed; you are defining your digital socioeconomic identity.