Abstract:Several approaches focus on how to automatically capture the latent features from original diffusion data and predict the future scale of cascades utilizing a black box framework. However, they ignore the penetrating insight into the underlying mechanism that how each participant is involved in the cascade. In this work, we bridge the gap between prediction and understanding of information diffusion by incorporating deep learning techniques and social psychology. To characterize individual participation driven by both subjective and objective impetus and integrate it into the macro-level cascade, we propose an end-to-end model, named PFDID, which is designed based on the field dynamics theory of psychology, including the intrinsic cognition field and the extrinsic environment field. We represent these two field dynamics respectively with the pairwise semantic relation between the message itself and corresponding comment and the forwarder’s micro-community activity embedding to provide educated explanations for forwarding behaviour. Afterwards, the cross infusion mechanism is designed to calculate the mutual influence of inhomogeneous field dynamics inside users and cross influence of homogeneous field dynamics among individuals, whose output is fed into the diffusion network aggregation layer for the cascade size prediction. Extensive experiments on two typical social networks, Sina Weibo and Twitter, manifest that the proposed PFDID outperforms state-of-the-art approaches. Our model achieves excellent prediction results, with MSLE = 1.856 on Sina Weibo and MSLE = 1.962 on Twitter, providing 6.54% and 10.53% relative performance gains, respectively. Furthermore, the interpretability is also discussed based on detailed visualization. We observe that the psychological impetus behind social behaviour varies mainly following two patterns with the spread of information, including gradual change and joint influence. Additionally, the indirect dependencies have also been verified.

DMHANT: DropMessage Hypergraph Attention Network for Information Propagation Prediction

Mining Subcascade Features for Cascade Outbreak Prediction in Big Networks.

MS-HGAT: Memory-Enhanced Sequential Hypergraph Attention Network for Information Diffusion Prediction

Hierarchical Attention Neural Network for Information Cascade Prediction

GRASS: Learning Spatial–Temporal Properties from Chainlike Cascade Data for Microscopic Diffusion Prediction

Information Cascade Prediction of complex networks based on Physics-informed Graph Convolutional Network

MCDAN: a Multi-scale Context-enhanced Dynamic Attention Network for Diffusion Prediction

Hypergraph Label Propagation Network.

H-Diffu: Hyperbolic Representations for Information Diffusion Prediction

Preference Enhanced Social Influence Modeling for Network-Aware Cascade Prediction

Anytime User Engagement Prediction in Information Cascades for Arbitrary Observation Periods

Continuous-Time Graph Learning for Cascade Popularity Prediction

Hierarchical Information Enhancement Network for Cascade Prediction in Social Networks

HierCas: Hierarchical Temporal Graph Attention Networks for Popularity Prediction in Information Cascades

A Cascading Diffusion Prediction Model in Micro-blog Based on Multi-dimensional Features.

Hierarchical Community-Level Information Diffusion Modeling in Social Networks

Improving Information Cascade Modeling by Social Topology and Dual Role User Dependency

Understanding information diffusion with psychological field dynamic

Topic-Aware Masked Attentive Network for Information Cascade Prediction

Path Prediction of Information Diffusion Based on a Topic-oriented Relationship Strength Network

DyHGCN: A Dynamic Heterogeneous Graph Convolutional Network to Learn Users' Dynamic Preferences for Information Diffusion Prediction