Modelling the Latent Semantics of Diffusion Sources in Information Cascade Prediction.

Area of Science:

• Information Science
• Computer Science
• Social Network Analysis

Background:

• Information cascade prediction is crucial for product recommendation and public opinion management.
• Existing models overlook the unique role of diffusion sources, treating them as ordinary users.
• Diffusion sources provide latent topic and pattern insights, vital for accurate cascade prediction.

Purpose of the Study:

• To propose a novel framework, Diffusion Source latent Semantics-Fused cascade prediction (DSSF), for information cascade prediction.
• To effectively model the implicit semantics of diffusion sources and their interaction with users.
• To address the long-tailed distribution of diffusion sources and improve prediction for less common sources.

Main Methods:

• Diffusion source embedding to represent the unique characteristics of initial users.
• A co-attention-based fusion gate to integrate diffusion source semantics with user embeddings.
• An adversarial training framework to transfer knowledge from prevalent (head) to rare (tail) diffusion sources.

Main Results:

• Modeling diffusion sources significantly enhances information cascade prediction performance.
• The proposed DSSF framework demonstrates superior prediction accuracy compared to baseline models.
• Adversarial training effectively improves prediction for cascades originating from tail sources, mitigating performance disparities.

Conclusions:

• Incorporating diffusion source semantics is critical for advancing information cascade prediction.
• The DSSF framework offers a robust approach to capture source-user interactions and improve prediction.
• Adversarial training is a viable strategy to handle data imbalance issues related to diffusion source distribution.