Recognising Personality Traits Using Facebook Status Updates

Golnoosh Farnadi, Susana Zoghbi, Marie‐Francine Moens, Martine De CockAAAI 2021 · DOI

TL;DR

This paper explores using machine learning to infer personality traits from Facebook status updates, demonstrating that even with limited data, models can outperform simple baselines. Crucially, training models on a different domain (essays) and applying them to Facebook data shows that personality recognition generalizes across social media platforms, opening doors for more robust and adaptable systems.

What Problem Does This Paper Address?

Understanding a user's personality is increasingly valuable for personalized online experiences, from targeted advertising to recommender systems. However, reliably inferring these traits from the vast and often noisy data generated on social media platforms like Facebook presents a significant challenge. Traditional methods might rely on explicit self-reporting, which can be biased or unavailable. This paper aims to bridge this gap by leveraging the rich, albeit unstructured, text of Facebook status updates, combined with social network and temporal information, to automatically predict personality traits.

Methods

The researchers employed a supervised machine learning approach. They trained binary classifiers for each of the Big Five personality traits (extraversion, neuroticism, agreeableness, conscientiousness, and openness to experience). The core of their methodology involved extracting a diverse set of features from Facebook data:

  • LIWC features: These are linguistic features derived from the Linguistic Inquiry and Word Count (LIWC) tool, capturing aspects like word counts, psychological processes (e.g., anger words), relativity, personal concerns, and linguistic dimensions.
  • Social Network features: These capture the user's network structure, including size, betweenness, density, and brokerage.
  • Time-related features: These analyze the temporal patterns of status updates, such as frequency per day and distribution across different time slots.
  • Other features: A miscellaneous set including total status count, capitalization, and specific string occurrences.

They compared three standard machine learning algorithms: Support Vector Machine (SVM) with a linear kernel, k-Nearest Neighbors (kNN), and Naive Bayes (NB). Performance was evaluated using a weighted average of precision, recall, and F-measure, with weights adjusted for class imbalance. A key aspect of their work was exploring "cross-domain learning," where models trained on one dataset (e.g., Facebook) were applied to another (e.g., essays), and vice versa, to assess generalization capabilities.

Datasets

The primary dataset used was a corpus of 250 Facebook users and 9917 status updates collected from the myPersonality project. Each user in this dataset had completed a questionnaire to establish their personality traits.

For cross-domain learning, they utilized a second corpus of 2468 essays written by psychology students, which were also labeled with Big Five personality traits. This essay corpus provided linguistic features but lacked social network and temporal information.

Key Insights and Ideas

  • Facebook status updates are rich personality indicators: Even with a relatively small number of status updates, the models could significantly outperform a majority class baseline, indicating that users' writing styles and content on Facebook reveal meaningful personality cues.
  • No single feature set reigns supreme: Different feature categories (linguistic, social, temporal) contributed differently to predicting various personality traits, suggesting that a multi-faceted approach is necessary for comprehensive personality recognition.
  • Cross-domain learning is powerful: The most striking finding is that models trained on essay data performed well when applied to Facebook data, and vice versa. This demonstrates that personality trait recognition models can generalize across different types of social media and text-based platforms. This is a significant advantage, as it allows for training more robust models by combining data from various sources or applying models to platforms where specific training data is scarce.
  • Social features correlate strongly with extraversion: The study found that social network features were significantly correlated with extraversion, aligning with previous research and suggesting that extroverts tend to have larger, sparser networks where their friends might not be closely connected.

Key Results

  • The models consistently outperformed the majority class baseline across various feature sets and algorithms. For instance, in precision, SVM with social features achieved 0.67 for extraversion, significantly better than the baseline.
  • When combining features, performance generally improved. For example, combining LIWC, Social, and Time features with SVM for Openness achieved an F-measure of 0.67.
  • In cross-domain learning experiments, training on both Facebook and essay data simultaneously often yielded the best results. For example, the "Essay+FB-FB" approach achieved an F-measure of 0.65 for extraversion.
  • The study also highlighted specific linguistic cues associated with traits, such as neurotic users being more likely to use anger words and agreeable users more likely to use sexual words.

Limitations and Future Work

The authors acknowledge the relatively small size of the Facebook corpus (250 users). They also note that for highly imbalanced classes, like Openness (where ~70% of users were labeled as "Open"), it was harder to improve upon the majority class baseline.

Future work could involve more fine-grained feature selection, as demonstrated by their example using only LIWC:anger features for neuroticism, which yielded better results than broader feature sets. Exploring more advanced feature engineering techniques and investigating the impact of user demographics and cultural differences on personality expression in online text are also promising avenues. The success of cross-domain learning suggests further research into transfer learning techniques for personality recognition across diverse online platforms.

Summary updated

Abstract

Gaining insight in a web user's personality is very valuable for applications that rely on personalisation, such as recommender systems and personalised advertising. In this paper we explore the use of machine learning techniques for inferring a user's personality traits from their Facebook status updates. Even with a small set of training examples we can outperform the majority class baseline algorithm. Furthermore, the results are improved by adding training examples from another source. This is an interesting result because it indicates that personality trait recognition generalises across social media platforms.
BibTeX
@inproceedings{farnadi2021recognising,
  author = {Golnoosh Farnadi and Susana Zoghbi and Marie‐Francine Moens and Martine De Cock},
  title = {{Recognising Personality Traits Using Facebook Status Updates}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v7i2.14470},
}
Figure