Multi-view Contrastive Learning for Event Reconstruction in Dissemination of information
Zhongchao He, Yuefeng Ma, Shumei Wang · 2024
The rapid development of social media provides people with abundant channels for information dissemination, but also leads to the problem of information fragmentation. Such fragmentation causes individuals to incompletely or incorrectly understand public events. To address the issue of event reconstruction, we propose a Multi-view Contrastive Event Reconstruction (MCER) model. The core idea is to maximize the information between different views of the same event and minimize the information between different events using contrastive learning, thereby recombining fragmented images in feature space to comprehensively reconstruct public events. MCER adopts a three-tower architecture comprising a momentum encoder and a weight-sharing encoder, and performs representation learning through a designed contrastive loss to achieve multi-view event reconstruction. Considering the lack of public social media multi-view datasets, we construct the Mul-View-data dataset. Results demonstrate MCER’s favorable performance over existing methods on both public datasets and Mul-View-data. Specifically, for multi-view contrastive learning, MCER significantly outperforms self-supervised and supervised methods by using our improved contrastive loss.