Research map / Graph learning / Graph self-supervised learning
DAG-Aware Variational Autoencoder for Social Propagation Graph Generation
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 DAVA, an attention-based DAG-aware exponential VAE that generates large-scale social propagation DAGs conditioned on ~1M crawled user attributes, and introduces a Credibility Erosion Effect decay.
Problem
Existing social propagation models are oversimplified, scenario-specific, and detached from real-world user attributes, and real propagation datasets are too small to evaluate downstream methods at scale.
Approach
Combines a GRU + exponential-distribution VAE for graph-level topology with a user-feature attention mechanism for edge-level generation, using BFS permutations ordered by user importance and a sliding window to scale generation, plus a decay on predecessors modeling credibility erosion.
Key results
- DAVA achieves MMD2 of 0.149 (Twitter15), 0.134 (Twitter16), 0.201 (Weibo), versus best baseline D-VAE at 0.271, 0.216, 0.255, a ~40% average reduction.
- DAVA generates graphs of 10,000 nodes in about 0.5 h, where baselines fail at that scale (e.g., GraphRNN and GVAE MM fail at 10,000 nodes).
- Augmenting source localization training with 1,000 DAVA-generated graphs raises GCNSI accuracy from 0.532 to 0.613 and TGASI from 0.787 to 0.825, above SOTA-generated augmentation and traditional simulation controls.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DAVA-generated propagation graphs to augment training data | GCNSI · Twitter source localization | accuracy: 0.532 -> 0.613 | Helped |
| use DAVA-generated propagation graphs to augment training data | TGASI · Twitter source localization | accuracy: 0.787 -> 0.825 | Helped |
| augment training with simulated data from traditional propagation models | GCNSI · Twitter source localization | accuracy: 0.532 -> 0.512 | Hurt |
| augment training with simulated data from traditional propagation models | TGASI · Twitter source localization | accuracy: 0.787 -> 0.755 | Hurt |
| use exponential rather than normal distribution for graph-level latent space (ablation) | DAVA · Twitter16 | MMD²: 0.207 -> 0.134 | Helped |
| use interpretable user-relationship attention rather than Transformer attention (ablation) | DAVA · Twitter16 | MMD²: 0.151 -> 0.134 | Helped |
| use interpretable user-relationship attention rather than an autoregressive generative model (ablation) | DAVA · Twitter16 | MMD²: 0.322 -> 0.134 | Helped |
| use the custom loss rather than binary cross-entropy loss (ablation) | DAVA · Twitter16 | MMD²: 0.236 -> 0.134 | Helped |
| add the CEE-based decay mechanism (ablation) | DAVA · Twitter16 | MMD²: 0.181 -> 0.134 | Helped |
Surprising findings
- Traditional simulated propagation data (SI, SIR, IC, LT) reduced downstream source localization performance below the no-augmentation original baseline.
- Rumors spread more virally and slower but persist longer, and most rumors involve fewer participants than non-rumor events.
Limitations
- Relies on crawled Twitter and Weibo data whose user attributes are not fully reflected in propagation datasets; generalization to other platforms is untested.
- Credibility Erosion Effect decay factor beta is chosen by hand and validated only empirically.
Experimental setup
Datasets: Twitter15 Twitter16 Weibo
Benchmarks: Twitter15 Twitter16 Weibo
Metrics: MMD2 CCDF of structural virality graph scale generation time source detection accuracy
Baselines: DAGG GVAE MM D-VAE GraphVAE GraphRNN GCNSI TGASI
Builds on: D-VAE GraphRNN Set2Set Transformer
Closest papers by meaning
- Non-Euclidean Mixture Model for Social Network Embedding (NeurIPS 2024 · similarity 0.86)
- DiffIM: Differentiable Influence Minimization with Surrogate Modeling and Continuous Relaxation (AAAI 2025 · similarity 0.85)
- REM: A Scalable Reinforced Multi-Expert Framework for Multiplex Influence Maximization (AAAI 2025 · similarity 0.85)
- When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors (NeurIPS 2026 · similarity 0.83)
- AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks (AAAI 2025 · similarity 0.83)
- Adversarial Socialbots Modeling Based on Structural Information Principles (AAAI 2024 · similarity 0.83)
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.