Multimodal Graph Negative Learning
Researchers have introduced GraphMNL, a novel framework for multimodal graph negative learning designed to address semantic imbalance in attributed graphs. This approach uses negative learning to guide learning across different semantic branches, such as text and images, rather than forcing imitation of a dominant branch. GraphMNL identifies dominant and inferior branches through reliability arbitration and applies target-preserving negative learning to suppress unlikely classes. Experiments show GraphMNL achieving top performance on the Grocery datasets with 72.47% accuracy and on the Reddit M datasets with a 76.60 F1 score. AI
IMPACT Introduces a new method for handling semantic imbalance in multimodal attributed graphs, potentially improving representation learning for complex relational systems.