Knowledge-Preserving Incremental Social Event Detection via Heterogeneous GNNs

Yuwei Cao, Hao Peng, Jia Wu, Yingtong Dou, Jianxin Li, Philip S. Yu

Introduction

Social media platforms generate a continuous stream of information, where identifying "events"—clusters of messages related to specific real-world occurrences—is vital for crisis management and recommendation systems. However, traditional event detection methods struggle with three primary challenges: the "moving target" problem where new event types emerge constantly, the structural complexity of social messages (involving users, hashtags, and entities), and the "memory gap" caused by catastrophic forgetting in incremental learning settings.

To address these issues, this paper introduces KPGNN (Knowledge-Preserving Incremental Heterogeneous Graph Neural Network), a framework designed to detect social events in real-time streams while maintaining learned knowledge over time.

Methodology: The KPGNN Framework

KPGNN leverages Heterogeneous Graph Neural Networks to preserve the rich context of social media interactions. The architecture is divided into several key components that transform raw messages into actionable event clusters.

The KPGNN architecture: from raw message preprocessing to GNN-based embedding and contrastive learning.

Heterogeneous Social Message Modeling

Rather than treating messages as isolated text, KPGNN constructs a Heterogeneous Information Network (HIN). This network links messages to their constituent words, named entities, and associated users. This HIN is subsequently mapped into a homogeneous message graph where edges represent shared social elements, ensuring that the underlying context is preserved during the embedding process.

Contrastive Learning for Dynamic Classes

Because the number of social events is unknown and constantly changing, KPGNN avoids standard cross-entropy loss. Instead, it employs a dual-loss strategy:

  1. Triplet Loss: Pushes messages from the same event closer in the embedding space while separating those from different events.
  2. Global-Local Pair Loss: Maximizes mutual information between individual message embeddings and the overall graph structure, allowing the model to distinguish events based on their unique social footprints.

Incremental Learning and Maintenance

KPGNN is designed for an inductive life-cycle, allowing it to handle streaming data without requiring full retraining from scratch. This process is managed through three distinct stages: Pre-training, Detection, and Maintenance.

The three-stage inductive life-cycle of KPGNN, illustrating how the model updates its knowledge and prunes obsolete data.

In the Maintenance stage, the model periodically resumes training with new data blocks to update its internal knowledge. To prevent the graph from growing indefinitely, the researchers implemented a pruning strategy. Interestingly, keeping only the latest message block for graph reconstruction yielded the best performance, suggesting that immediate temporal context is more critical for social media learning than maintaining a long, stale history.

Experimental Evaluation

The model was evaluated on two large-scale benchmarks: a Twitter dataset (68,841 tweets, 503 events) and MAVEN (a Wikipedia-based event dataset). KPGNN significantly outperformed baselines such as BERT and specialized models like EventX and PP-GCN.

  • Accuracy: In incremental settings, KPGNN achieved a 16% gain in Normalized Mutual Information (NMI) and a 319% gain in Adjusted Mutual Information (AMI) over the EventX baseline.
  • Efficiency: By utilizing the "latest message" maintenance strategy, the model maintained high performance while significantly reducing GPU memory usage and training time compared to "keep-all" history strategies.

Conclusion

KPGNN represents a significant advancement in streaming social event detection by combining the inductive power of GNNs with contrastive learning and intelligent graph maintenance. While the model excels at detection, future work may explore causal discovery between events and tracking how social clusters evolve and morph over time.

Summary updated

Abstract

Social events provide valuable insights into group social behaviors and public concerns and therefore have many applications in fields such as product recommendation and crisis management. The complexity and streaming nature of social messages make it appealing to address social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns. Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge. In this paper, we propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. To acquire more knowledge, KPGNN models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction. To continuously adapt to the incoming data, KPGNN adopts contrastive loss terms that cope with a changing number of event classes. It also leverages the inductive learning ability of GNNs to efficiently detect events and extends its knowledge from previously unseen data. To deal with large social streams, KPGNN adopts a mini-batch subgraph sampling strategy for scalable training, and periodically removes obsolete data to maintain a dynamic embedding space. KPGNN requires no feature engineering and has few hyperparameters to tune. Extensive experiment results demonstrate the superiority of KPGNN over various baselines.
BibTeX
@inproceedings{cao2021knowledge,
  author = {Yuwei Cao and Hao Peng and Jia Wu and Yingtong Dou and Jianxin Li and Philip S. Yu},
  title = {{Knowledge-Preserving Incremental Social Event Detection via Heterogeneous GNNs}},
  booktitle = {The Web Conference},
  year = {2021},
  doi = {10.1145/3442381.3449834},
}
Figure