Self-Supervised Hypergraph Representation Learning for Sociological Analysis

Xiangguo Sun, Hong Cheng, Bo Liu, Jia Li, Guandong Xu, Hongzhi Yin

TL;DR

This paper introduces a self-supervised framework that bridges the gap between "black-box" data mining and human-centric sociological theory. By representing social environments as hypergraphs and their interactions as line graphs, the authors successfully quantify subjective concepts like social conformity and polarization to improve tasks like personality detection and fake news identification.

What Problem Does This Paper Address?

Traditional data mining for online behavior analysis (like recommendation systems) focuses on finding patterns in latent variable spaces. While effective, these models often lack "human-understandable" logic and ignore decades of sociological research.

Sociologists have long studied concepts like social equivalence (similar people inhabit similar environments) and social conformity (individuals align with group norms). However, these are notoriously difficult to measure in massive Online Social Networks (OSNs) because they are subjective and computationally expensive to model. The authors aim to create a "task-free" methodology that can learn these complex social structures automatically and apply them to various downstream data mining tasks.

Methods

The authors propose a multi-layered approach centered on Hypergraphs:

  1. Social Environment Awareness: Instead of simple pair-wise links (User A follows User B), they use hyperedges to represent "environments" (groups of users sharing attributes and connections). They use a negative log-likelihood target to learn which users belong in which hyperedge.
  2. Fast Line Graph Construction: To model how different environments influence each other, they build a "line graph" where each node is a hyperedge. They introduce a novel, efficient algorithm using random walks to estimate similarities between environments, avoiding the memory bottlenecks of traditional methods.
  3. Hypergraph Neural Network (HNN): They design a neural network to model "influence flow" in four directions: user-to-user, user-to-environment, environment-to-user, and environment-to-environment.
  4. Dual & Contrastive Learning: To make the model "task-free," they use a self-supervised approach. The model learns by trying to reconstruct the original user features from the complex hypergraph representations. This allows the learned "embeddings" to be "plugged" into any task (like rating prediction) without starting from scratch.

Datasets

The framework was evaluated on five diverse real-world datasets:

  • MovieLens: Movie ratings (user-item interactions).
  • Facebook: Social network following relations.
  • FakeNews: Retweeting networks (identifying fake vs. real news).
  • PersonalityCafe: An online forum where users have Myers-Briggs (MBTI) labels.
  • KarateClub: A small, classic network used for structural visualization.

Key Insights and Ideas

  • Beyond Pair-wise Relations: Simple graphs are too restrictive. Hypergraphs allow a single "edge" to contain many nodes, perfectly capturing the sociological concept of a "social environment."
  • The Power of the Line Graph: Modeling the relationship between groups (environments) is just as important as modeling individuals. The line graph allows the model to capture how social trends jump from one community to another.
  • Quantifying the Subjective: The authors provide mathematical definitions for sociological terms. For example, they define Social Polarization using "group entropy"—measuring the uncertainty/confidence of group members' beliefs.
  • "Pluggable" Intelligence: Because the model is self-supervised, it learns a deep understanding of social structure first. This pre-trained knowledge consistently improves performance on specific tasks compared to models trained only on those tasks.

Key Results

  • Personality Detection: On the PersonalityCafe dataset, the model achieved an accuracy of 0.4555, significantly outperforming baselines like DeepInf (0.2937) and metapath2vec (0.1856).
  • Fake News Identification: The model reached an AUC of 0.9507, proving that social structure and "conformity" patterns are strong signals for detecting misinformation.
  • Rating Prediction: In MovieLens tests, the model achieved the lowest Mean Absolute Error (MAE) of 0.1397, beating specialized collaborative filtering models like NCF.
  • Qualitative Success: t-SNE visualizations showed that the model creates much clearer clusters of user types compared to standard AutoEncoders or Graph Convolutional Networks (GCNs).

Limitations and Future Work

While the paper introduces a "fast" line graph algorithm, hypergraph methods are inherently more computationally demanding than simple graph methods. The authors also note that sociological theory is a vast field; this paper focuses on a few core criteria (conformity, equivalence, polarization), but many other theories (like social exchange or conflict theory) remain to be integrated.

Future work could involve exploring more dynamic hypergraphs that evolve in real-time as users join or leave online communities.

Summary updated

Abstract

Modern sociology has profoundly uncovered many convincing social criteria for behavioral analysis. Unfortunately, many of them are too subjective to be measured and very challenging to be presented in online social networks (OSNs) for the large data volume and complicated environments to be explored. On the other hand, data mining techniques can better find data patterns but many of them leave behind unnatural understanding to humans. Although there are some works trying to integrate social observations for specific tasks, they are still hard to be applied to more general cases. In this paper, we propose a fundamental methodology to support the further fusion of data mining techniques and sociological behavioral criteria. Our highlights are three-fold: First, we propose an effective hypergraph awareness and a fast line graph construction framework. The hypergraph can more profoundly indicate the interactions between individuals and their environments because each edge in the hypergraph (a.k.a hyperedge) contains more than two nodes, which is perfect to describe social. A line graph treats each social environment as a super node with the underlying influence between different environments. In this way, we go beyond traditional pair-wise relations and explore richer patterns under various sociological criteria; Second, we propose a novel hypergraph-based neural network to learn social influence flowing from users to users, users to environments, environment to users, and environments to environments. The neural network can be learned via a task-free method, making our model very flexible to support various data mining tasks and sociological analysis; Third, we propose both qualitative and quantitive solutions to effectively evaluate the most common sociological criteria like social conformity, social equivalence, environmental evolving and social polarization. Our extensive experiments show that our framework can better support both data mining tasks for online user behaviors and sociological analysis.
BibTeX
@article{sun2023self,
  author = {Xiangguo Sun and Hong Cheng and Bo Liu and Jia Li and Hongyang Chen and Guandong Xu and Hongzhi Yin},
  title = {{Self-Supervised Hypergraph Representation Learning for Sociological Analysis}},
  journal = {IEEE Trans. Knowl. Data Eng.},
  year = {2023},
  doi = {10.1109/tkde.2023.3235312},
}
Figure