Research map / Graph learning / Graph self-supervised learning
Towards Effective, Efficient and Unsupervised Social Event Detection in the Hyperbolic Space
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 HyperSED, an unsupervised social event detection framework that builds a semantic anchor message graph and learns hyperbolic graph and differentiable structural information partitioning trees, achieving competitive accuracy and up to 37.41x speedup over HISEvent.
Problem
Social event detection on large, complex, and dynamic social message streams is hard: existing methods use all messages (resource-intensive), Euclidean embeddings (insufficient for hierarchical structure), and often require the number of events or supervision.
Approach
Group semantically similar messages into anchor nodes to form a Semantic-based Anchor Message Graph (SAMG), embed anchors with a hyperbolic graph autoencoder on the Poincaré ball, and optimize a differentiable structural information partitioning tree (with the root at the origin) to read out events without a preset event count.
Key results
- Offline, HyperSED reaches NMI 0.84 / AMI 0.77 / ARI 0.44 on English Twitter and NMI 0.68 / AMI 0.62 / ARI 0.62 on French Twitter, the best in the table on both datasets.
- Efficiency speed-up over HISEvent reaches up to 37.41x on French Twitter online blocks and at least 12.10x (e.g., 12.10x in M2), with 65 s vs 25,372 s running time offline.
- Reported average improvements over prior SOTA are about 2% NMI, 2% AMI and 25% ARI for incremental SED per the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperSED instead of the strongest baselines | Unsupervised offline SED · English Twitter offline | ARI: +0.27 | Helped |
| Use HyperSED instead of the strongest baselines | Unsupervised offline SED · French Twitter offline | NMI: +0.02 | Helped |
| Use HyperSED instead of HISEvent | Unsupervised SED · French Twitter offline | overall time: 27.70 times | Helped |
| Use HyperSED instead of HISEvent | Unsupervised SED · Twitter English and French online blocks | NMI | Mixed |
| Add the Semantic-based Anchor Message Graph (SAMG) (ablation) | HyperSED · Twitter English and French | clustering performance | Helped |
| Add differentiable structural information (DSI) (ablation) | HyperSED · Twitter English and French | clustering performance | Helped |
| Add attribute edges (E_a) (ablation) | HyperSED · Twitter English and French | clustering performance | Helped |
| Add semantic edges (E_s) (ablation) | HyperSED · Twitter English and French | clustering performance | Helped |
| Vary structural entropy height H across 2, 3, and 4 (ablation) | HyperSED · English Twitter online | clustering performance | No clear effect |
| Vary anchor count using ε in {10, 20, 30} (ablation) | HyperSED · English Twitter online | clustering performance | No clear effect |
Surprising findings
- Online results show marginal decreases in several message blocks (e.g., M2, M3, M5), which the authors attribute to error accumulation during anchor construction.
- Basic SBERT with k-means (given ground-truth event count) is competitive with many graph methods on NMI/AMI in several blocks.
Limitations
- Performance slightly decreases in some online blocks due to error accumulation in anchor construction.
- Anchor grouping can incorrectly merge semantically similar but distinct-event messages, a trade-off for efficiency.
Experimental setup
Datasets: English Twitter (68,841 messages, 503 events) French Twitter (64,516 messages, 257 events)
Benchmarks: English Twitter French Twitter
Metrics: NMI AMI ARI running time speed-up
Baselines: BERT SBERT EventX KPGNN QSGNN HISEvent
Builds on: HISEvent Differentiable Structural Information hyperbolic graph convolution SBERT
Closest papers by meaning
- Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection (AAAI 2024 · similarity 0.89)
- THGNets: Constrained Temporal Hypergraphs and Graph Neural Networks in Hyperbolic Space for Information Diffusion Prediction (AAAI 2025 · similarity 0.85)
- Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning Approach for Rumor Detection (AAAI 2024 · similarity 0.82)
- MSR: A Multifaceted Self-Retrieval Framework for Microscopic Cascade Prediction (AAAI 2025 · similarity 0.82)
- THGFormer: Time-Aware Hypergraph Learning for Multimodal Social Media Popularity Prediction (Student Abstract) (AAAI 2024 · similarity 0.82)
- Hypergraph Representation Learning with Hyperlink Random Effects (NeurIPS 2026 · 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.