Facebook versus Twitter: Differences in Self-Disclosure and Trait Prediction

Kokil Jaidka, Sharath Chandra Guntuku, Lyle H. UngarICWSM 2018

Introduction

Computational social science research often treats social media behavior as a monolith, yet users intuitively adapt their self-presentation to suit different platforms. This study addresses the "cross-platform gap" by examining how the same individuals disclose information differently on Facebook versus Twitter. The researchers aimed to determine which platform serves as a better "window" into a user’s psychological and demographic traits and whether predictive models can generalize across these distinct digital environments.

Methodology

The researchers employed a multi-pronged natural language processing (NLP) approach to analyze user language across both platforms.

Study methodology overview including feature extraction and predictive modeling.

The experimental pipeline involved:

  • Feature Extraction: Standard techniques including n-grams, LIWC (psychological themes), LDA Topics (unsupervised word clusters), and Word2Vec embeddings.
  • Predictive Modeling: Elastic-Net regression for continuous traits (age, stress, empathy) and Random Forest classifiers for binary traits (gender).
  • Ablation Studies: To control for Facebook’s higher text volume, researchers conducted experiments matching total word counts (WL), post counts (PL), and words-per-post (PW) between the two datasets.

Dataset

The study utilized a unique dataset of 523 users who maintained active accounts on both platforms and completed psychological surveys. The corpus included approximately 647,000 Facebook posts and 862,000 tweets. Ground truth was established via self-reported age, gender, stress (Cohen’s Stress scale), and empathy (Interpersonal Reactivity Index).

Linguistic and Psychological Insights

Analysis revealed a distinct "Honesty vs. Ambition" divide in platform usage. Users tend to use Facebook for "honest" self-representation—discussing family, personal concerns, and deep emotions. Conversely, Twitter is used to "think aloud," focusing on professional drives, cognitive processes, and ambitions.

Linguistic clusters showing differences between platforms (Twitter vs. Facebook) and age groups (Younger vs. Older).

Key linguistic markers identified include:

  • Facebook: Dominated by words related to family, home, and religious affiliations. It appears to be a more socially "safe" space for users to vent or seek support, as stress correlates more strongly with negative emotion here than on Twitter.
  • Twitter: Characterized by "netspeak" (rt, lol, u), second-person pronouns, and cognitive process words (comparing, differentiating).

Key Results

Facebook models consistently outperformed Twitter models in predictive accuracy across the board:

  • Age Prediction: Facebook achieved a Mean Absolute Error (MAE) of 5.71 years, significantly better than Twitter’s 8.08 years.
  • Gender Prediction: Facebook reached an accuracy of 0.91, while Twitter peaked at 0.80.
  • The Equalization Effect: In ablation studies, restricting Facebook’s language sample to match Twitter’s volume narrowed the gap. For age prediction, the platforms became roughly equal, though Facebook remained superior for gender prediction even with limited data.
  • The Transfer Penalty: Models do not generalize well across platforms. For instance, an empathy model trained on Facebook performs significantly worse on Twitter, highlighting that psychological traits manifest differently depending on the digital context.

Conclusion and Limitations

The findings suggest that Facebook language is a more potent predictor of personal traits, partly due to higher disclosure volume and partly due to the nature of the content shared. However, the study is limited by a primarily US-based, white demographic and potential self-selection bias. Future work may expand this analysis to include visual data and metadata to create a more holistic digital phenotype.

Summary updated

BibTeX
@inproceedings{jaidka2018facebook,
  author = {Kokil Jaidka and Sharath Chandra Guntuku and Lyle H. Ungar},
  title = {{Facebook versus Twitter: Differences in Self-Disclosure and Trait Prediction}},
  booktitle = {ICWSM},
  year = {2018},
}
Figure