Research map / Label-efficient and robust learning / Test-time adaptation
Ghidorah: Towards Robust Multi-Scale Information Diffusion Prediction via Test-Time Training
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 Ghidorah, a test-time training model for multi-scale information diffusion prediction that adapts a shared encoder per test cascade with an MAE-style auxiliary loss, plus invariant learning and a gradient-consistency loss.
Problem
Information diffusion prediction (macroscopic cascade size and microscopic next-user prediction) assumes IID train/test data, but social cascades suffer out-of-distribution shifts from user behavior variability, leaving static trained models brittle at inference.
Approach
Applies test-time training to diffusion prediction: for each test cascade, fine-tunes the shared feature extractor with a masked-autoencoder self-supervised loss, aligns SSL gradients with the main task gradients via a consistency loss, and uses variance risk extrapolation over augmented environments for invariant learning.
Key results
- Microscopic prediction: Hits@100 on Christianity reaches 0.8059 vs 0.6394 for the best baseline CE-GCN.
- Macroscopic prediction: MSLE on Christianity is 0.356 vs 0.384 for SFA, a relative reduction of over 15% over the best baseline.
- Ablation: removing TTT raises Christianity MSLE from 0.356 to 0.485; removing the consistency loss raises it to 0.372.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ghidorah instead of the strongest baseline | Microscopic prediction · Christianity, Android, Memetracker, and Douban | Hits@100: 0.6394 -> 0.8059; 0.2766 -> 0.3214; 0.6509 -> 0.7276; 0.4275 -> 0.4845 | Helped |
| Use Ghidorah instead of the strongest baseline | Microscopic prediction · Christianity, Android, Memetracker, and Douban | MAP@100: 0.2054 -> 0.2673; 0.0727 -> 0.0841; 0.2256 -> 0.2521; 0.1260 -> 0.1417 | Helped |
| Use Ghidorah instead of the strongest baseline | Macroscopic prediction · Christianity, Android, Memetracker, and Douban | MSLE: 0.384 -> 0.356; 0.130 -> 0.127; 0.438 -> 0.414; 0.326 -> 0.290 | Helped |
| Add HGNN to capture global user dynamic interactions (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.1965 -> 0.2673; 0.1092 -> 0.1417 | Helped |
| Add test-time training (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.2214 -> 0.2673; 0.1286 -> 0.1417 | Helped |
| Add consistency loss to align auxiliary and primary task updates (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.2622 -> 0.2673; 0.1379 -> 0.1417 | Helped |
| Add invariant learning (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.2586 -> 0.2673; 0.1326 -> 0.1417 | Helped |
| Use MAE rather than contrastive learning as the auxiliary task (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.2302 -> 0.2673; 0.1254 -> 0.1417 | Helped |
| Use MAE rather than BYOL as the auxiliary task (ablation) | Ghidorah · Christianity and Douban | MAP@100: 0.2471 -> 0.2673; 0.1295 -> 0.1417 | Helped |
| Vary the MAE mask ratio; best performance is reported at 0.4 (ablation) | Ghidorah · Christianity | MAP@100 | Mixed |
| Vary auxiliary-task loss weight α; best performance is reported at 0.5 (ablation) | Ghidorah · Christianity | MAP@100 | Mixed |
| Vary the number of constructed environments; best performance is reported at 6 (ablation) | Ghidorah · Christianity | MAP@100 | Mixed |
Limitations
- Unconstrained test-time training can degrade main-task performance, so the consistency loss is needed.
- Performance is sensitive to hyperparameters such as gradient steps (overfitting beyond 30 steps) and number of environments.
Experimental setup
Datasets: Christianity Android Memetracker Douban
Benchmarks: Christianity Android Memetracker Douban
Metrics: MSLE Hits@k MAP@k
Baselines: DeepCas DeepHawkes CasCN CasFlow TopoLSTM NDM Inf-VAE DyHGCN MS-HGAT CE-GCN FOREST DMT-LIC
Builds on: test-time training (Sun et al. 2020) masked autoencoder variance risk extrapolation GCN HGNN
Closest papers by meaning
- Enhancing Multi-Scale Diffusion Prediction via Sequential Hypergraphs and Adversarial Learning (AAAI 2024 · similarity 0.89)
- MSR: A Multifaceted Self-Retrieval Framework for Microscopic Cascade Prediction (AAAI 2025 · similarity 0.85)
- Learning Divergence Fields for Shift-Robust Graph Representations (ICML 2024 · similarity 0.85)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.84)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.84)
- Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis (ICLR 2025 · similarity 0.84)
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.