The Secret Lives of Names?: Name Embeddings from Social Media

Junting Ye, Steven SkienaKDD 2019 · DOI

TL;DR

This paper introduces publicly available name embeddings trained on Twitter data, demonstrating they perform comparably to private email-based embeddings for tasks like gender and ethnicity prediction. Crucially, these Twitter embeddings significantly improve lifespan prediction models, revealing latent socioeconomic distinctions previously hidden in demographic data.

What Problem Does This Paper Address?

Names are rich with information, often revealing gender, ethnicity, religion, and even socioeconomic status. Traditional methods for extracting this information from names often rely on explicit features like substrings, which can be brittle. Name embeddings, a form of distributed representation, have shown promise in capturing these nuanced associations. However, prior work on name embeddings was trained on private email data and thus not accessible to the broader research community. This limits their widespread adoption and application. The authors aim to bridge this gap by creating publicly available name embeddings from social media (Twitter) and demonstrating their utility in a novel application: lifespan prediction.

Methods

The core of the paper is the creation and evaluation of name embeddings derived from Twitter data. They explore nine different training corpora based on Twitter user interactions (retweets, mentions, followers, followees) to find the most effective. The embeddings are learned using a skip-gram model with negative sampling, a standard technique in word embedding generation.

To evaluate their effectiveness, they use these embeddings for:

  1. Gender, Ethnicity, and Nationality Classification: Comparing Twitter embeddings against email embeddings and a baseline of random embeddings.
  2. Lifespan Modeling: Integrating name embeddings as features into linear regression models that predict lifespan, alongside traditional demographic features (birth year, state, gender, ethnicity, nationality). They compare models with and without name embeddings, and with shuffled embeddings as a control.
  3. Latent Factor Analysis: Using residual analysis on the lifespan models to identify specific name characteristics (e.g., diminutive vs. formal names, ethnic origins) that contribute positively or negatively to lifespan.

Datasets

  • Twitter Data: Collected from Twitter's API, including:
  • Tweets (retweets and mentions)
    • Follower and followee lists from a random set of "seed users" (selected to have a moderate number of connections and not be celebrities or bots).
  • Email Data: Used as a baseline, from a previous study [37] trained on 57 million email users.
  • U.S. Census Data: Used for ground truth labels for gender (1990 Census) and ethnicity (2000 Census).
  • Social Security Death Index (SSDI): A dataset of over 85 million death records, used for lifespan prediction experiments. This dataset contains full name, date of birth, and date of death.

Key Insights and Ideas

  • Twitter Embeddings are Viable: Name embeddings trained on public Twitter data can achieve performance comparable to those trained on private email data for tasks like gender and ethnicity prediction, even with a smaller training corpus.
  • Mentions Capture Stronger Gender Homophily: Embeddings derived from Twitter "mentions" show the highest performance in gender recognition, suggesting users are more likely to interact with and mention others of the same gender.
  • Removing Celebrities Improves Followee Embeddings: Followee lists, which often contain celebrities, yield less homophilous embeddings. Removing celebrity names from these lists ("Followee*") strengthens the homophily signal.
  • Aggregated Embeddings Shine: Combining various Twitter interaction types (Retweet, Mention, Follower, Followee) into an "Aggregated" embedding generally yields the best performance among Twitter-derived embeddings.
  • Name Embeddings Unlock Latent Socioeconomic Factors: Integrating name embeddings into lifespan prediction models significantly improves accuracy. Residual analysis reveals that these embeddings capture fine-grained distinctions, such as socioeconomic status, indicated by preferences for formal over diminutive names and the positive association of certain ethnic origins (e.g., Ashkenazi Jewish) with longer lifespans.

Key Results

  • Gender Classification: Mention embeddings achieved the highest gender classification accuracy, outperforming other Twitter embeddings and Email embeddings.
  • Ethnicity Classification: Aggregated Twitter embeddings achieved comparable performance to Email embeddings, with Follower and Followee embeddings showing strong results.
  • Lifespan Prediction:
    • Adding name embeddings (both Twitter and Email) to demographic features consistently improved lifespan prediction models, with all p-values < 0.01.
  • The improvement was attributed to the latent signals within the embeddings, not just increased feature dimensionality.
  • Twitter embeddings performed similarly to Email embeddings in lifespan modeling.
  • Birth year was the most powerful single predictor, yielding an average error of 8.052 years.
  • Latent Factors:
    • Diminutive first names (e.g., Wm, Gus) were associated with negative lifespan gains, while their formal counterparts (e.g., William, Daniel) showed positive gains, suggesting socioeconomic distinctions.
  • A significant portion of the most favorable last names (17 out of 20) had Ashkenazi Jewish origins, highlighting the embeddings' ability to capture group-specific longevity patterns.

Limitations and Future Work

  • Public Release: The authors have released their Twitter name embeddings to the research community via www.name-prism.com.
  • Latent Factor Interpretation: While the paper demonstrates the predictive power of name embeddings and identifies some latent factors, the exact nature of all hidden properties encoded within them remains an "exciting open question for further research." The authors speculate these include subtle socioeconomic status and fine-grained group distinctions.
  • Future Applications: The authors believe these embeddings can be valuable for strengthening predictive models in various social, economic, and medical science domains, especially in data-scarce situations where names serve as proxies for demographic attributes.

Summary updated

BibTeX
@inproceedings{ye2019secret,
  author = {Junting Ye and Steven Skiena},
  title = {{The Secret Lives of Names?: Name Embeddings from Social Media}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  doi = {10.1145/3292500.3330751},
}
Figure