Research map / Image generation / Fast diffusion sampling

CasFT: Future Trend Modeling for Information Popularity Prediction with Dynamic Cues-Driven Diffusion Models

AAAI 2025 · Image generation › Fast diffusion sampling · approach: Diffusion models as perception tools · 8 citations (OpenAlex)

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 CasFT, which uses neural ODE-derived dynamic cues and a conditional diffusion model to generate future popularity-increase trends, then fuses them with observed cascade spatiotemporal features to predict information popularity.

Problem

Predicting future incremental popularity of information cascades on social platforms, where existing methods model only the observed period and ignore the uncertain popularity trend between observation and prediction times.

Approach

Use neural ODEs to model the continuous growth rate and cumulative popularity after observation, then use these dynamic cues as conditions for a DDIM-based diffusion model that generates segmented future popularity increments, which are concatenated with cascade representations for the final prediction.

Key results

• CasFT reports MSLE 3.8546 on Twitter (1 day) vs 4.7799 for the best baseline CasFlow (19.36% improvement).
• On APS (3 years) CasFT achieves MSLE 1.2468 vs 1.4370 for CasFlow (13.24% improvement).
• On Weibo (0.5 hours) CasFT achieves MSLE 2.1728 vs 2.3370 for CasFlow (7.02% improvement); ablation removing future trend modeling worsens MSLE by up to 20.75%.

Measured effects reported in the paper

Surprising findings

• The CasFT-FM variant (MLP replacing the future trend module) is worse than CasFT but still outperforms CasFT-w/o FT, suggesting the trend block's benefit is not solely from extra predicted sequences.
• Euler ODE solver performs worst, while dopri5 gives the best results across datasets.

Limitations

• Improvements on MAPE are small on some datasets (e.g., 2.2% on Twitter 1 day), and the paper's reported gains are sensitive to diffusion steps on Twitter.
• Future growth rate relies on autoregressive ODE prediction, and the ODE state is constrained by the initial state, which the authors note can bias estimates.

Experimental setup

Datasets: Twitter APS Sina Weibo

Benchmarks: Twitter APS Sina Weibo

Metrics: MSLE MAPE

Baselines: Feature-based SEISMIC DeepCas DeepHawkes CasCN VaCas CasFlow CTCP

Builds on: neural ODE denoising diffusion implicit models DeepHawkes CasFlow

Closest papers by meaning

• Public Opinion Field Effect and Hawkes Process Join Hands for Information Popularity Prediction (AAAI 2025 · similarity 0.88)
• SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity (CVPR 2025 · similarity 0.85)
• Improving Multimodal Social Media Popularity Prediction via Selective Retrieval Knowledge Augmentation (AAAI 2025 · similarity 0.83)
• Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.83)
• Enhancing Multi-Scale Diffusion Prediction via Sequential Hypergraphs and Adversarial Learning (AAAI 2024 · similarity 0.83)
• THGFormer: Time-Aware Hypergraph Learning for Multimodal Social Media Popularity Prediction (Student Abstract) (AAAI 2024 · similarity 0.82)