[Paper Review] Detecting the Age of Twitter Users.

Benjamin Paul Chamberlain, Clive Humby+1 more|arXiv (Cornell University)|Jan 18, 2016

Bayesian Modeling and Causal Inference1 references3 citations

TL;DR

This paper proposes a language-independent method to detect Twitter users' ages using social network structure—specifically, whom they follow—rather than relying on names or linguistic features. By applying a Bayesian classification model to 700 million accounts, it achieves an F1 score of 0.86, enabling accurate age inference without requiring user-provided birth data or language-specific models.

ABSTRACT

Twitter provides an extremely rich and open source of data for studying human behaviour at scale. It has been used to advance our understanding of social network structure, the viral flow of information and how new ideas develop. Enriching Twitter with demographic information would permit more precise science and better generalisation to the real world. The only demographic indicators associated with a Twitter account are the free text name, location and description fields. We show how the age of most Twitter accounts can be inferred with high accuracy using the structure of the social graph. Besides classical social science applications, there are obvious privacy and child protection implications to this discovery. Previous work on Twitter age detection has focussed on either user-name or linguistic features of tweets. A shortcoming of the user-name approach is that it requires real names (Twitter names are often false) and census data from each user's (unknown) birth country. Problems with linguistic approaches are that most Twitter users do not tweet (the median number of Tweets is 4) and a different model must be learnt for each language. To address these issues, we devise a language-independent methodology for determining the age of Twitter users from data that is native to the Twitter ecosystem. Roughly 150,000 Twitter users specify an age in their free text description field. We generalize this to the entire Twitter network by showing that age can be predicted based on what or whom they follow. We adopt a Bayesian classification paradigm, which offers a consistent framework for handling uncertainty in our data, e.g., inaccurate age descriptions or spurious edges in the graph. Working within this paradigm we have successfully applied age detection to 700 million Twitter accounts with an F1 Score of 0.86.

Motivation & Objective

• To address the lack of reliable demographic data in Twitter by enabling accurate age detection without relying on user-provided names or linguistic content.
• To overcome limitations of prior approaches that depend on real names (often false) or language-specific models, which are impractical for multilingual or low-tweet users.
• To develop a scalable, generalizable method for inferring age from native Twitter data—specifically, social graph relationships—across diverse populations.
• To provide a consistent framework for handling uncertainty in age descriptions and noisy social graph edges using Bayesian inference.
• To enable more precise social science research and improve child protection mechanisms by accurately estimating user age at scale.

Proposed method

• Leverages the social graph—specifically, the set of users a given account follows—as the primary feature for age prediction, avoiding reliance on user names or tweet content.
• Applies a Bayesian classification framework to model uncertainty in age descriptions and potential spurious edges in the social network, enabling robust inference.
• Trains on a seed set of ~150,000 users who self-report their age in their profile description to generalize age patterns across the broader network.
• Uses a probabilistic model to infer age groups based on the age distribution of users in a user’s social circle, assuming age-related homophily in following behavior.
• Scales the method to 700 million Twitter accounts using efficient inference techniques suitable for large-scale graphs.
• Employs a multi-class classification approach with age groups (e.g., 18–24, 25–34, etc.) to balance accuracy and generalization across diverse populations.

Experimental results

Research questions

• RQ1 Can user age be accurately predicted using only social network structure, without linguistic or name-based features?
• RQ2 How does the performance of age detection vary across different age groups when relying solely on who a user follows?
• RQ3 To what extent can self-reported age in profiles be used to train a model that generalizes to the broader Twitter population?
• RQ4 How does the Bayesian framework handle uncertainty in age labels and noisy social graph data?
• RQ5 Can this method be applied at scale to millions or billions of users without language-specific tuning?

Key findings

• The method achieves an F1 score of 0.86 when predicting age groups across 700 million Twitter accounts, demonstrating high accuracy at scale.
• The approach is language-independent, eliminating the need for language-specific models or linguistic analysis of tweets.
• It outperforms prior methods that rely on user names or tweet content, especially for users who post few or no tweets (median of 4 tweets).
• The Bayesian framework effectively handles uncertainty in self-reported ages and spurious social connections, improving robustness.
• The model generalizes well from a small seed set of 150,000 users with self-reported ages to the entire Twitter network.
• The study demonstrates that social graph structure alone contains sufficient signal for reliable age inference, even without direct linguistic or name-based cues.