Abstract
Traditional data mining often treats social interactions as simple pairwise links, overlooking the complex, group-based structures defined by sociological theory. This paper proposes a self-supervised framework that utilizes hypergraphs and line graphs to model "social environments" and their mutual influences. By quantifying subjective concepts such as social conformity and equivalence, the model learns robust user representations in a task-free manner. Experimental results across diverse datasets—including personality detection and fake news identification—demonstrate that this sociologically-grounded approach significantly outperforms state-of-the-art baselines.
Introduction
Online Social Networks (OSNs) provide a wealth of data for behavioral analysis, yet most current models rely on "black-box" latent variable patterns that lack human-understandable logic. Sociological research has long identified key mechanisms of human behavior, such as social equivalence (where individuals in similar environments exhibit similar behaviors) and social conformity (the tendency to align with group norms). However, translating these subjective, high-level concepts into computational models for massive datasets remains a significant challenge.
The authors address this gap by moving beyond simple graph structures. While traditional graphs represent pairwise relations (e.g., User A follows User B), they fail to capture the nuances of a "social environment"—a group of users sharing attributes, interests, and connections. By leveraging hypergraphs, where a single edge can connect multiple nodes, the proposed framework can more accurately represent these complex social structures.
Methodology: A Dual-Learning Framework
The core of the proposed method is a self-supervised, multi-layered architecture designed to learn task-agnostic social representations.
1. Social Environment Awareness
The model first identifies "environments" using hyperedges. Instead of predefined groups, it learns user-to-hyperedge memberships by optimizing a negative log-likelihood target. This allows the model to discover latent communities based on shared interaction patterns.
2. Line Graph Construction and Influence Flow
To model how different environments interact, the authors construct a "line graph" where each node represents a hyperedge. A novel, random-walk-based algorithm efficiently estimates similarities between these environments, overcoming the memory constraints of traditional line graph methods. A Hypergraph Neural Network (HNN) then models the flow of influence across four dimensions: user-to-user, user-to-environment, environment-to-user, and environment-to-environment.
3. Task-Free Self-Supervised Learning
To ensure the learned embeddings are "pluggable" into various downstream tasks, the framework employs a contrastive learning objective. It attempts to reconstruct original user features from the complex hypergraph representations. This "Social Reconstruction" module ensures the model captures the essence of social equivalence and conformity without being tied to a specific label during training.
Experimental Results
The framework was evaluated on five real-world datasets: MovieLens (ratings), Facebook (social links), FakeNews (retweets), PersonalityCafe (MBTI labels), and KarateClub (structure).
Quantitative Performance
The model showed remarkable gains in specialized tasks:
- Personality Detection: Achieved 0.4555 accuracy on PersonalityCafe, nearly doubling the performance of baselines like metapath2vec (0.1856).
- Fake News Identification: Reached an AUC of 0.9507, suggesting that social conformity patterns are highly predictive of misinformation spread.
- Rating Prediction: Outperformed specialized collaborative filtering models on MovieLens with a Mean Absolute Error (MAE) of 0.1397.
Qualitative Analysis
The effectiveness of the representation learning is most evident in visual cluster analysis.
As shown in the t-SNE visualizations, the proposed model produces significantly clearer and more distinct clusters of user types compared to standard AutoEncoders or Graph Convolutional Networks (GCNs). This indicates that the sociologically-grounded hypergraph approach captures structural nuances that traditional methods miss.
Conclusion
By bridging the gap between data mining and sociological theory, this research provides a powerful tool for understanding complex online behaviors. The use of hypergraphs to quantify subjective concepts like conformity and polarization allows for a more "human-centric" approach to AI. While hypergraph methods are computationally intensive, the introduction of efficient line graph algorithms makes this framework viable for large-scale OSNs. Future work will look toward dynamic hypergraphs that can adapt to the real-time evolution of online communities.
Summary updated