Research map / Image generation / Fast diffusion sampling
Enhancing Multi-Scale Diffusion Prediction via Sequential Hypergraphs and Adversarial Learning
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 MINDS, a unified model for macroscopic cascade size and microscopic next-user prediction that combines sequential diffusion hypergraphs, social-homophily GCN, and a shared LSTM regularized by adversarial and orthogonality losses.
Problem
Information diffusion prediction at both macro (cascade size) and micro (next influenced user) levels has mostly been studied separately; prior unified models ignore cross-cascade global interactions and feature contamination between shared and task-specific representations.
Approach
Models cross-cascade interactions via sequential hypergraphs over time windows (HGNN with a fusion layer), learns social homophily with GCN, and extracts task-shared features with a shared LSTM kept pure via adversarial training and orthogonality constraints.
Key results
- On microscopic prediction, MINDS reaches Hits@100 of 0.6250 on Christianity versus 0.5923 for DyHGCN, the strongest baseline there.
- On macroscopic prediction, MINDS achieves MSLE of 0.572 on Christianity, 0.151 on Android, 0.404 on Douban and 0.506 on Memetracker, the lowest among all methods in Table 4.
- Ablation shows removing both adversarial and orthogonality losses (w/o AdvDiff) degrades performance, e.g., Christianity MSLE rises from 0.572 to 0.971.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Christianity | Hits@100: 0.5923 -> 0.6250 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Christianity | MAP@100: 0.1746 -> 0.2054 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Christianity | MSLE: 1.692 -> 0.572 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Android | Hits@100: 0.2596 -> 0.2766 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Android | MAP@100: 0.0675 -> 0.0727 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Android | MSLE: 0.201 -> 0.151 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Douban | Hits@100: 0.3329 -> 0.3641 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Douban | MAP@100: 0.0897 -> 0.1213 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Douban | MSLE: 0.465 -> 0.404 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Memetracker | Hits@100: 0.5710 -> 0.5790 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Memetracker | MAP@100: 0.1595 -> 0.1638 | Helped |
| Use MINDS instead of the strongest reported baseline | multi-scale diffusion prediction · Memetracker | MSLE: 0.535 -> 0.506 | Helped |
Surprising findings
- Removing the macro task (w/o Macro) sharply worsens micro metrics on Christianity but mainly harms MSLE dramatically (9.255 vs 0.572), indicating strong reliance on the joint objective.
Limitations
- Ablation results on only two datasets in the main text; the paper itself notes this may be insufficient.
- Performance gains on sparse datasets like Christianity and Douban are less consistent across ablation variants.
- Hyperparameters were selected by inspection of test-time trends rather than a stated validation protocol, and the hardware is unspecified.
Experimental setup
Datasets: Christianity Android Douban Memetracker
Benchmarks: Christianity Android Douban Memetracker
Metrics: MSLE MAP@k Hits@k
Baselines: DeepCas DeepHawkes CasCN CasFlow TCSE-net TopoLSTM NDM SNIDSA Inf-VAE DyHGCN TAN-DURD FOREST
Builds on: HGNN FOREST DMT-LIC GCN LSTM
Closest papers by meaning
- THGNets: Constrained Temporal Hypergraphs and Graph Neural Networks in Hyperbolic Space for Information Diffusion Prediction (AAAI 2025 · similarity 0.90)
- Ghidorah: Towards Robust Multi-Scale Information Diffusion Prediction via Test-Time Training (AAAI 2025 · similarity 0.89)
- Decoupling User Relationships Guides Information Diffusion Prediction (Student Abstract) (AAAI 2024 · similarity 0.88)
- A Generalized Neural Diffusion Framework on Graphs (AAAI 2024 · similarity 0.87)
- Towards a Mechanistic Explanation of Diffusion Model Generalization (ICML 2025 · similarity 0.86)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.86)
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.