Prefix-Conditioned Diffusion for Early Rumor Detection Under Sparse and Incomplete Social Cascades

Haowei Xu, Chao Gao†, Xianghua Li, Zhen Wang(† corresponding author)

ACM Transactions on Intelligent Systems and Technology (TIST) 2026 Journal

Early rumor detection is hardest exactly when intervention is most valuable: the observed cascade is still sparse, yet the decision cannot wait for a mature propagation trace. A short prefix can admit several structurally different continuations, while standard detectors compress that prefix into one representation and one point prediction. We introduce RumorDiffusion, a class-conditioned topology diffusion framework that makes this missing-state ambiguity available to the decision layer. Conditioned on the same observed prefix and fixed semantic features, the model generates a potential representation for each veracity hypothesis. A Transformer arbiter compares these class potentials, and repeated reverse trajectories provide a dispersion signal for validation-calibrated prediction or deferral. Our analysis decomposes the resulting decision error into class-potential approximation, finite-trajectory sampling, and reject cost, linking each term to an observable diagnostic. Experiments on Twitter15, Twitter16, Weibo, and PHEME are organized around five research questions covering point discrimination, evidence accumulation, mechanism attribution, selective computation, and source-only transfer. Dataset-specialized models set the point-prediction frontier; RumorDiffusion exposes additional calibration, stability, coverage, and latency behavior through its class-potential interface. The reported sampling sweep selects a single reverse trajectory as the strongest quality–latency operating point, while the source-only study places RumorDiffusion in the lower-ECE region for both evaluated transfer pairs. These results establish structured continuation modeling as a practical decision interface for determining when an incomplete cascade contains sufficient evidence for action.

NeuronCtrl: Geometry-Aware Safe Closed-Loop Generative Control for Neuronal Microenvironment Dynamics

Haowei Xu, Yixin Chen, Wanyi Fu, Hongbin Han, Zhaoheng Xie

International Conference on Machine Learning (ICML) 2026 Spotlight

TL;DR: We introduce NEURONCTRL, a geometry-aware framework for safe closed-loop generative control of neuronal microenvironment dynamics using morphology-aware neural operators and flow-matching, with barrier-based safety mechanisms and real-time performance.

Subphenotype Identification for Sepsis-Associated Acute Kidney Injury using Graph Bidirectional Mamba Networks

Haowei Xu, Wentie Liu, Tongyue Shi, Guilan Kong

IEEE Journal of Biomedical and Health Informatics 2025 Journal

We introduce a bidirectional Mamba-based graph learning framework that uncovers clinically meaningful subphenotypes for sepsis-associated acute kidney injury by modeling multimodal EHR dynamics and cross-patient relationships.

Neural Granger Causal Discovery for Derangements in ICU-Acquired Acute Kidney Injury Patients

Haowei Xu, Wentie Liu, Tongyue Shi, Guilan Kong

AMIA Annual Symposium Proceedings 2025 Conference

This work discovers causal pathways driving acute kidney injury in the ICU by pairing neural Granger causality with temporal multimodal measurements, enabling actionable risk stratification for critical care teams.

[Code]

SIGRL: Sociologically-Informed Graph Representation Learning for Social Influence Prediction

Haowei Xu, Chao Gao, Xingyu Li, Zhihai Wang

IEEE Transactions on Network Science and Engineering 2025 Journal

SIGRL integrates sociological priors into graph representation learning to model influence propagation with improved fidelity across complex social interaction networks.

Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference

Mingxin Li, Yuchen Zhang, Haowei Xu, Xingyu Li, Chao Gao, Zhihai Wang

AAAI 2025 Conference

The model infers latent social structures that fuse multimodal signals, advancing fake news detection with structured reasoning across text, vision, and interaction cues.

[DOI]

D2: Customizing Two-Stage Graph Neural Networks for Early Rumor Detection through Cascade Diffusion Prediction

Haowei Xu, Chao Gao, Xingyu Li, Zhihai Wang

WSDM 2025 Conference

We design a diffusion-aware dual-stage graph neural network that anticipates rumor spread patterns and flags misinformation during its early cascade evolution across social platforms.

HyperIDP: Customizing Temporal Hypergraph Neural Networks for Multi-Scale Information Diffusion Prediction

Haowei Xu, Chao Gao, Xingyu Li, Zhihai Wang

COLING 2025 Conference

HyperIDP leverages temporal hypergraph neural modeling to capture multi-scale diffusion signals across heterogeneous communities, improving forecasting accuracy for evolving narratives.

RumorMixer: Exploring Echo Chamber Effect and Platform Heterogeneity for Rumor Detection

Haowei Xu, Chao Gao, Xingyu Li, Zhihai Wang

ECML-PKDD 2024 Conference

RumorMixer disentangles echo chambers and cross-platform heterogeneity through a multi-view graph design, enabling robust rumor detection in diverse social media ecosystems.