Multi-modal advertising visual information fusion using dynamic heterogeneous graph neural networks

To enhance the fusion performance of multimodal advertising content in visual comprehension and behaviour prediction, this paper proposes a DHGNN-based multimodal information fusion model.The model comprises modal feature embedding, dynamic node/edge attribute modelling, improved heterogeneous graph convolutions, and a multi-task output layer, optimising CTR, CVR, and semantic classification.Compared with the conventional Hetero GCN and MM Attn models, the proposed DHGNN improved the AUC by 3.8-5.9p.p. on CTR, and the F1-score was 0.726, which was 8.8 percentage points above Homogeneous GCN.Moreover, the inference latency is reduced to 19.4 ms, while maintaining competitive predictive accuracy.Across Ali-AD and Tencent-Ad360, DHGNN achieves AUC gains of 3.8-5.9percentage points on CTR and reaches an F1-score of 0.726 on semantic classification, providing quantitative evidence that the proposed dynamic heterogeneous graph modelling improves multimodal fusion under heterogeneous and time-varying advertising data for multimodal recommendation systems.