Graph attention contrastive learning with missing modality for multimodal recommendation

Wenqian Zhao

doi:10.1016/j.knosys.2025.113035

Graph attention contrastive learning with missing modality for multimodal recommendation

2025-02-01

0

PRE

AI

DOI：10.1016/j.knosys.2025.113035

Original

Original request for help

Save

Abstract

En 中文

Multimodal recommendation plays an important role in many online content-sharing platforms. Most existing reported approaches of multimodal recommendation employ user-interaction graphs or auxiliary graphs (e.g., user-user or item-item relation graphs) to augment user and/or item representations. However, real-world data suffer the problem of missing modality which affects recommendation performance. In this paper, we propose the Graph Attention Contrastive Learning with Missing Modality (MMGACL) model utilizing modality complementation and modality fusion of modality-aware user-item graphs to enhance recommendations. In particular, we construct user-item bipartite graphs for each modality and extract item subgraphs, leveraging contextual information to enhance item representations. Thereafter, we employ a bimodal attention mechanism to provide complementary information across modalities and fuse different modalities. The fused item representations are combined with user-item interactions to complement user information. Finally, we perform graph contrastive learning on the completed global graph to maximize mutual information between users and items and learn more accurate embedding representations. Extensive experiments on four benchmark datasets demonstrate the effective performance of our proposed model versus several state-of-the-art methods in scenarios with missing modality.

Keywords:

Multimodal recommendation

Missing modality

Contrastive learning

Graph neural network

Graph attention contrastive learning with missing modality for multimodal recommendation

Abstract

Journal

Organization