Data Augmented Graph Neural Networks for Personality Detection

Yangfu Zhu, Yue Xia, Meiling Li, Tingting Zhang, Bin WuAAAI 2024 · DOI

Personality detection—the task of identifying "Big Five" traits (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) from social media text—is a critical area of research with applications ranging from personalized marketing to mental health monitoring. However, it is a "data-hungry" task facing a chronic "data drought." Obtaining ground-truth labels typically requires participants to complete lengthy psychological inventories, resulting in datasets that are significantly smaller than standard NLP benchmarks.

To address this, researchers presented Semi-PerGCN at AAAI-24. This semi-supervised framework leverages a heterogeneous graph structure and consistency-based data augmentation on unlabeled posts to achieve state-of-the-art results across multiple benchmarks.

The Semi-PerGCN Framework

Semi-PerGCN operates by integrating domain-specific psychological knowledge directly into its architecture. Instead of treating social media posts as simple sequences of words, the model constructs a heterogeneous graph for each user. This graph includes three distinct types of nodes:

  1. User Nodes: Representing the individual whose personality is being predicted.
  2. Word Nodes: Representing the specific vocabulary used in the user's posts.
  3. LIWC Category Nodes: Utilizing the Linguistic Inquiry and Word Count (LIWC) lexicon, words are linked to psychological categories (e.g., "sad" and "crying" both link to a "Negative Emotion" node).
The Semi-PerGCN architecture, showing (A) supervised personality detection via GCN and (B) unsupervised consistency learning on augmented graphs.

This structure allows a Graph Convolutional Network (GCN) to learn high-level psycholinguistic relationships. A path-specific attention mechanism then helps the model weigh which psychological paths are most relevant for a specific user’s personality traits.

Leveraging Unlabeled Data via Consistency Learning

A core innovation of Semi-PerGCN is its ability to utilize massive amounts of unlabeled social media data through Consistency Learning. The model takes an unlabeled user's posts and generates "augmented" versions using two primary techniques:

  • LIWC Synonym Replacement (SR): Swapping a word with another from the same psychological category. This preserves the psychological "signal" while varying the literal text.
  • Random Deletion (RD): Removing non-psychological words to force the model to focus on the most meaningful content.
Illustration of data augmentation strategies: Synonym Replacement within LIWC categories and Random Deletion of non-psychological nodes.

The model is trained to ensure its personality predictions for the original post and the noisy augmented post are nearly identical. This consistency constraint acts as a powerful regularizer, preventing the model from overfitting on small labeled sets and encouraging it to learn more robust, generalized representations.

Performance and Key Insights

The model was evaluated on three gold-standard datasets: Youtube (404 users), PAN2015 (294 users), and MyPersonality (10,000 users). Semi-PerGCN consistently outperformed specialized graph models like TrigNet and D-DGCN, achieving a 1.51% to 3.49% improvement in average Mean Absolute Error (MAE).

Key takeaways from the study include:

  • Psychology as a Structural Prior: By including LIWC categories as nodes, domain knowledge is baked into the architecture, reducing the need for the model to "guess" which words are psychologically relevant.
  • Low-Resource Resilience: The model showed particular strength in scenarios with very few labels, proving that unsupervised consistency learning effectively extends the utility of limited data.
  • Semantic vs. Psychological Noise: The augmentation strategy targets "psychological synonyms," making the model focus on underlying intent rather than specific word choice.

Future Directions

While Semi-PerGCN marks a significant step forward, the authors note that performance gains from adding unlabeled data eventually plateau. Future work may explore advanced metric learning, such as contrastive learning, to explicitly push different personality types apart in the embedding space. Additionally, addressing category imbalances in real-world social media data remains a priority for improving model reliability across diverse populations.

Summary updated

BibTeX
@inproceedings{zhu2024data,
  author = {Yangfu Zhu and Yue Xia and Meiling Li and Tingting Zhang and Bin Wu},
  title = {{Data Augmented Graph Neural Networks for Personality Detection}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i1.27823},
}
Figure