IEGSCL: Interaction-Enhanced Graph Neural Sequence Contrastive Learning for Microscopic Diffusion Prediction

Area of Science:

• Social Network Analysis
• Information Diffusion Dynamics
• Machine Learning

Background:

• Understanding user relationships and preferences is key to explaining information diffusion in social networks.
• Existing graph neural network (GNN) methods for information diffusion prediction often overlook crucial spread interaction feedback.
• Over-reliance on limited labeled data hinders the self-learning and generalization capabilities of current models.

Purpose of the Study:

• To propose a novel microscopic diffusion prediction model that effectively utilizes interaction feedback and unlabeled data.
• To enhance the learning of user representations by integrating social trust and interaction data.
• To improve the self-learning and generalization capabilities of information diffusion prediction models.

Main Methods:

• Developed an interaction-enhanced graph neural sequence contrastive learning (IEGSCL) model.
• Constructed a triple graph incorporating trust and interaction to capture diverse user relationships and preferences.
• Implemented a self-supervised graph contrastive learning module for user intent transfer and feature extraction from unlabeled data.
• Designed an information-driven gating strategy to adaptively integrate social and interactive intents into cascade modeling.
• Utilized maximum mean discrepancy (MMD) to align global relationship representations with local cascade encodings.

Main Results:

• The proposed IEGSCL model demonstrated superior performance compared to existing baseline methods.
• Experiments on four public datasets validated the effectiveness of the interaction-enhanced approach.
• The model successfully leveraged unlabeled data and interaction feedback for improved diffusion prediction.

Conclusions:

• The IEGSCL model offers a significant advancement in microscopic information diffusion prediction.
• Integrating interaction feedback and utilizing unlabeled data are crucial for enhancing model performance and generalization.
• The proposed methods provide a robust framework for understanding and predicting information spread in social networks.