Information Diffusion Prediction with Temporal and Structural Contrastive Learning Augmented Graph Neural Network

ACL ARR 2025 February Submission6887 Authors

Abstract:

Information diffusion prediction aims to forecast how information propagates through social networks. Current works have explored users' preferences from dynamic diffusion structures and social relations.
Despite recent advances, they generally share two natural deficiencies.
First, they generally fail to identify users' critical preferences hidden in noisy and complex user structures.
In addition, existing works primarily extract users' dynamic preferences within localized sub-graph structures, struggling to filter relevant preferences for the current cascade.
Thus, we propose \textbf{T}emporal and \textbf{S}tructural \textbf{C}ontrastive \textbf{L}earning \textbf{A}ugmented Graph Neural Network (\textbf{TSCLA}).
Specifically, we split the diffusion process into discrete periods and introduce a temporal contrastive learning module to extract users' {diversified} preferences across the diffusion process.
Furthermore, we introduce a hierarchical adaptation module that dynamically filters relevant preferences in each diffusion period.
In addition, we construct a heterogeneous graph to extend users' preferences and design a structural contrastive learning module for discerning critical user relations from noisy connections.
Experimental results on four real-world datasets demonstrate the superior performance of our model compared to state-of-the-art baselines.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: human behavior analysis

Contribution Types: Model analysis & interpretability

Languages Studied: None

Submission Number: 6887