Research map / Graph learning / Graph self-supervised learning
Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HISEvent, an unsupervised social event detector that builds message graphs with incremental 1D structural entropy minimization and partitions them with hierarchical 2D structural entropy minimization, without labels or a preset event count.
Problem
Social event detection with GNN-based methods misses semantically related message correlations, requires labeled data for training, and needs the number of events specified in advance.
Approach
Supplements attribute-based message graphs with semantic kNN edges chosen by incremental 1D structural entropy minimization, then partitions the graph with a hierarchical (subgraph-wise) greedy 2D structural entropy minimization that is much faster than the vanilla algorithm.
Key results
- Closed-set Event2012: ARI 0.50 and AMI 0.81 versus 0.22 and 0.73 for the strongest baseline (SBERT/KPGNN/QSGNN, using ground-truth event counts).
- Closed-set Event2018: ARI 0.44 (vs 0.16 for QSGNN) and AMI 0.66 (vs 0.62 for SBERT), a 175% ARI gain over SBERT.
- Hierarchical 2D SE minimization reduces running time by more than 97% on the largest Event2012 block M1 versus the vanilla algorithm, which takes over 5 days, with comparable performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic-similarity-based edges (ablation) | HISEvent · Event2012 closed-set | ARI: 0.24 -> 0.50 | Helped |
| add semantic-similarity-based edges (ablation) | HISEvent · Event2012 closed-set | AMI: 0.58 -> 0.81 | Helped |
| add semantic-similarity-based edges (ablation) | HISEvent · Event2012 open-set average | ARI: 0.40 -> 0.63 | Helped |
| add common-attribute-based edges (ablation) | HISEvent · Event2012 closed-set | AMI: 0.80 -> 0.81 | Helped |
| use SBERT instead of BERT for edge weights (ablation) | HISEvent · Event2012 closed-set | ARI: 0.25 -> 0.50 | Helped |
| replace vanilla 2D SE minimization with hierarchical 2D SE minimization (ablation) | HISEvent · Event2012 open-set average | AMI: 0.82 -> 0.82 | No clear effect |
| replace vanilla 2D SE minimization with hierarchical 2D SE minimization (ablation) | HISEvent · Event2012 | runtime: > 10 days | Helped |
| increase sub-graph size n from 100 to 400 (ablation) | HISEvent · Event2012 closed-set | ARI: 10% | Helped |
Surprising findings
- Plain SBERT clustering with ground-truth event counts outperforms the GNN-based KPGNN and QSGNN, indicating embedding quality matters more than the graph model.
- Using BERT rather than SBERT embeddings for edge weights and selection lowers HISEvent's performance, showing embeddings must reflect semantic similarity.
Limitations
- Performance depends on the quality of PLM embeddings reflecting semantic similarity; BERT embeddings degrade results.
- Batched/retrospective design; streaming extension is only discussed, not evaluated.
- Multilingual coverage limited to languages supported by SBERT models.
Experimental setup
Datasets: Event2012 Event2018
Benchmarks: Event2012 Event2018
Metrics: ARI AMI NMI
Baselines: KPGNN QSGNN EventX BERT SBERT HISEvent-vanilla
Builds on: structural entropy KPGNN QSGNN EventX SBERT
Closest papers by meaning
- Towards Effective, Efficient and Unsupervised Social Event Detection in the Hyperbolic Space (AAAI 2025 · similarity 0.89)
- Dependency Structure-Enhanced Graph Attention Networks for Event Detection (AAAI 2024 · similarity 0.86)
- Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection (AAAI 2024 · similarity 0.84)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.83)
- TGLsta: Low-resource Textual Graph Learning with Semantic and Topological Awareness via LLMs (AAAI 2025 · similarity 0.82)
- GLBench: A Comprehensive Benchmark for Graph with Large Language Models (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.