TL;DR
This paper presents a language-independent method to infer the age of Twitter users by analyzing who they follow, leveraging the idea that interests are indicative of age. Their Bayesian approach scales to 700 million users, offering a valuable demographic enrichment for social network analysis.
What Problem Does This Paper Address?
Social media platforms like Twitter are rich sources of data for understanding human behavior. However, a significant limitation for researchers is the scarcity of demographic information, particularly age. This lack of data hinders studies on social network structures, information diffusion, and the spread of trends. Existing methods often rely on linguistic analysis of tweets, which are language-dependent and struggle with users who tweet infrequently. This paper aims to overcome these limitations by developing a scalable, language-independent method to infer Twitter users' ages.
Methods
The core of the proposed method is a probabilistic Bayesian framework. The central hypothesis is that a user's interests, as revealed by who they follow on Twitter, are indicative of their age. The authors collect ground-truth age labels from a subset of Twitter users who explicitly state their age in their profile descriptions. They then use a Bayesian approach to generalize this information to the entire Twitter network.
Key methodological steps include:
- Data Collection and Ground-Truth Labeling: Crawling Twitter profiles to extract user descriptions and identifying age information using regular expressions (REGEX) across multiple languages. They filtered for active accounts to mitigate stale data.
- Feature Engineering: The features for age inference are the Twitter accounts that users follow. They identified accounts followed by a significant number of labeled users as potential features.
- Bayesian Inference: A Bayesian classification paradigm is employed to handle the inherent uncertainties (noisy labels, features, and survey estimates). The model uses a Naive Bayes assumption, where the decision to follow an account is independent given the user's age. They use Beta priors for Bernoulli distributions to model the likelihood of following a feature account.
- Scalability: The approach is designed to be massively scalable, allowing for inference on hundreds of millions of Twitter accounts.
Datasets
- Ground-Truth Data: Approximately 133,000 active Twitter users who disclosed their age in their profile descriptions. This data was used to train and evaluate the model.
- Augmented Ground-Truth: To improve performance in older age categories, the authors augmented their training data by including users labeled as "retired" (176,748) and "grandparents" (63,895).
- Large-Scale Twitter Data: The model was applied to infer the ages of 700 million Twitter accounts.
- Publicly Available Dataset: The authors released a public dataset containing an anonymized sample of their data, including a sparse bipartite adjacency matrix of user-feature relationships and age category labels.
Key Insights and Ideas
- Following as a Proxy for Interests: The fundamental insight is that "who you follow" is a powerful, language-independent indicator of your interests, which in turn can be linked to age. This bypasses the limitations of tweet content analysis.
- Bayesian Framework for Uncertainty: The use of a Bayesian framework is crucial for handling the inherent noise and uncertainty in social media data, leading to more robust age predictions.
- Scalability is Achievable: The paper demonstrates that it's possible to perform accurate age inference on a massive scale (700 million users) without requiring extensive manual feature engineering or language-specific models.
- Homophily Principle in Action: The work directly leverages the homophily principle, suggesting that individuals with similar attributes (like age) tend to form similar social ties and share similar interests.
Key Results
- The proposed probabilistic model can infer the age of Twitter users with high accuracy, scaling to 700 million accounts.
- The model performs comparably to state-of-the-art language-based models for age inference, while being language-independent and more scalable.
- For a three-class age model (under 18, 18-44, 45+), the approach achieved a Micro F1 score of 0.86, matching the Dutch Language Model (DLM) but with a more automated and scalable feature generation process.
- The model significantly outperforms an approach based solely on survey and census data, demonstrating its ability to adapt to the actual data observed on Twitter.
- The paper identified discriminative features (popular Twitter accounts) for different age groups, showing that younger users are characterized by interests in internet celebrities and gaming, while older users show interests in religious topics, sports, and politics.
Limitations and Future Work
- Data Sparsity and Noise: While the Bayesian approach helps, the inherent sparsity and noise in social media data remain challenges. The paper acknowledges that some accounts, even if labeled correctly, might not represent genuine human interests (e.g., bots).
- Ground-Truth Limitations: The ground-truth data is derived from self-reported ages, which can be inaccurate or outdated. The authors also note that their initial REGEX matching for age in descriptions could lead to false positives.
- Bias in Younger Age Categories: The model's performance is noted to be weaker for the 18-45 age category, potentially due to limitations in training data for this group.
- Future Work: The authors suggest that a hybrid approach combining their method with linguistic features could further improve performance. They also leave open the possibility of exploring other social network attributes beyond just "following" for age inference.
Summary updated