Content-aware susceptibility estimation for influence prediction on social networks

Predicting how information propagates in social networks is critical for applications such as recommendation, virality forecasting, and opinion management. While prior studies have explored susceptibility estimation, which aims to predict the probability that each individual will be influenced by an activated node (or seed node), the heterogeneity of the propagated content has been overlooked. In real-world scenarios, users exhibit varied responses to different types of information, even when disseminated by the same seed. In this paper, we formulate the Content-aware Susceptibility Estimation (CaSE) problem, which focuses on estimating peer-to-peer influence probabilities conditioned on multi-modal content. We propose DeepCS, an end-to-end deep learning framework that integrates network structures, multi-modal semantic features, and behavioral logs to capture fine-grained, peer-to-peer susceptibility. DeepCS employs a hierarchical cross-modal attention mechanism and a multi-path preference fusion strategy to model user interests and structure, mitigating semantic inconsistency and behavioral sparsity. It enhances user interest representation through contrastive learning and introduces a content-sensitive scoring mechanism to accurately predict influence sources and directional susceptibility. Comprehensive experiments on four real-world social networks demonstrate that DeepCS outperforms state-of-the-art approaches to estimating susceptibility.