Improving Information Diffusion Prediction by Tackling Noise and Sparsity Challenges

TL;DR

This paper introduces DDiff, a novel framework that improves information diffusion prediction by addressing noise and sparsity through denoising diffusion and cross-domain contrastive learning, leading to superior performance.

Contribution

The paper presents a new framework combining denoising diffusion and contrastive learning to effectively handle noise and data sparsity in IDP tasks, which is a novel approach.

Findings

DDiff outperforms existing methods in IDP accuracy.

The framework effectively reduces noise impact on diffusion modeling.

Cross-domain contrastive learning enhances knowledge transfer between domains.

Abstract

With the widespread use of online social media platforms, information diffusion has become a prevalent phenomenon, making Information Diffusion Prediction (IDP) increasingly important for various applications. Despite significant advancements in IDP research, existing methods often overlook issues of noise and sparsity in information diffusion data. User behaviors are frequently influenced by external factors, introducing noise into the data and hindering models' understanding of true diffusion patterns. Additionally, many users have limited interaction data, leading to data sparsity and restricting models' ability to effectively capture user preferences. To address these challenges, we propose a novel framework called DDiff, which tackles noise and sparsity issues through denoising diffusion and cross-domain contrastive learning. First, we introduce a graph learning encoder module that captures the social homophily of users through their relationships and higher-order connections via information diffusion hypergraphs (IDH). Next, a cross-domain contrastive learning module is designed to facilitate effective knowledge transfer between the information and social domains, addressing the sparsity problem. Furthermore, we propose a denoising diffusion module with IDH to effectively mitigate noise issues by introducing random noise in the forward process and iteratively recovering the corrupted embeddings in the reverse process. Finally, we implement a prediction module to determine the likelihood of subsequent users becoming infected. Experimental results demonstrate that DDiff significantly outperforms state-of-the-art methods in the information diffusion prediction task.