Inferring Human Traits From Facebook Statuses

Andrew Cutler, Brian Kulis

Abstract: This paper explores the use of language models to predict 20 human traits from users' Facebook status updates. The data was collected by the myPersonality project, and includes user statuses along with their personality, gender, political identification, religion, race, satisfaction with life, IQ, self-disclosure, fair-mindedness, and belief in astrology. A single interpretable model meets state of the art results for well-studied tasks such as predicting gender and personality; and sets the standard on other traits such as IQ, sensational interests, political identity, and satisfaction with life. Additionally, highly weighted words are published for each trait. These lists are valuable for creating hypotheses about human behavior, as well as for understanding what information a model is extracting. Using performance and extracted features we analyze models built on social media. The real world problems we explore include gendered classification bias and Cambridge Analytica's use of psychographic models.

Introduction

Digital footprints—specifically the unstructured text of social media posts—reveal deep-seated psychological and demographic traits. While previous research has often relied on structured data like Facebook "Likes," this study focuses on the rich linguistic content of status updates. Using the myPersonality dataset, which includes data from over 7 million users who participated in psychological quizzes between 2008 and 2012, the researchers investigated the extent to which "psychographic" information (personality, values, interests) can be extracted from everyday language. The goal was not only to achieve high predictive accuracy but also to address the "black box" problem by analyzing what these models actually learn.

Methods

The authors primarily utilized a Bag-of-Words (BoW) approach combined with tf-idf preprocessing. For continuous traits, they employed $\ell_2$ regularized regression (Ridge Regression), while categorical data was handled through a "one-vs-one" majority vote system.

In a comparison for gender prediction, the simple BoW model surprisingly outperformed a complex 49-layer character-level Convolutional Neural Network (char-CNN). To mitigate potential biases, such as models using gendered language as a proxy for other traits, the authors proposed a novel debiasing technique. This method explicitly encodes a protected attribute (like gender) during training but sets it to "unknown" during testing, forcing the model to identify predictive features independent of that attribute.

Key Results

The study demonstrates that simple linear models with high-dimensional BoW features are often more accurate and interpretable than complex neural networks for social media text.

Continuous Traits and IQ

For continuous traits, performance was measured using Explained Variance (EV). The model accounted for 12.8% of the variance in IQ scores, a significant result for text-only analysis. It also showed strong performance across the "Big Five" personality traits and various sensational interests.

Performance metrics (Explained Variance) for continuous traits including personality, sensational interests, and IQ.

Categorical Classification

Categorical traits such as gender, race, and political identity were also successfully classified. Gender prediction reached an accuracy of 92.8%, significantly outperforming human baselines. The model also reached 76.4% accuracy on binary political classification (Liberal vs. Conservative) and successfully expanded this to a 12-way classification across various ideologies.

Classification results for categorical traits including gender, race, and political identity.

Feature Analysis and Bias

Analysis of feature weights provided insights into human behavior. High-IQ users tended to use academic language (e.g., "physics"), while lower-IQ scores were associated with relationship-oriented words (e.g., "family"). However, the authors also found that models can rely on proxies; for instance, "militarism" was often predicted by "gamer" language like "xbox." Their debiasing method successfully reduced reliance on gendered words in religious classification without sacrificing overall accuracy.

Conclusion

This research underscores the predictive power of simple, interpretable models in psychographic profiling. While status text alone is highly effective, future work could incorporate social network structures or utilize transfer learning to improve predictions for traits with limited data samples. The findings serve as both a benchmark for trait inference and a cautionary analysis of the biases inherent in social media-based modeling.

Summary updated

Abstract

This paper explores the use of language models to predict 20 human traits from users' Facebook status updates. The data was collected by the myPersonality project, and includes user statuses along with their personality, gender, political identification, religion, race, satisfaction with life, IQ, self-disclosure, fair-mindedness, and belief in astrology. A single interpretable model meets state of the art results for well-studied tasks such as predicting gender and personality; and sets the standard on other traits such as IQ, sensational interests, political identity, and satisfaction with life. Additionally, highly weighted words are published for each trait. These lists are valuable for creating hypotheses about human behavior, as well as for understanding what information a model is extracting. Using performance and extracted features we analyze models built on social media. The real world problems we explore include gendered classification bias and Cambridge Analytica's use of psychographic models.
BibTeX
@inproceedings{cutler2018inferring,
  author = {Andrew Cutler and Brian Kulis},
  title = {{Inferring Human Traits From Facebook Statuses}},
  booktitle = {SocInfo},
  year = {2018},
  url = {https://openreview.net/forum?id=Y8sgmBuKQCJ},
}
Figure