GAT++: Adaptive Relation-Aware Graph Attention Networks

ICLR 2026 Conference Submission22076 Authors

Keywords: Novel View Synthesis, Dynamic Scene, Gaussian Splatting

Abstract: Leveraging high-order structural semantics in knowledge graphs (KGs) is critical for modeling complex user preferences in recommendations. However, during multi-hop propagation, semantic noise arising from heterogeneous relation distributions obscures meaningful preferences, making it challenging to learn robust user-item representations. To address this challenge, we propose GAT++, a novel graph convolutional network that integrates relation-aware attention mechanisms with contrastive denoising regularization to learn robust and expressive user-item representations. At its core, GAT++ introduces an adaptive attention module that captures multiple semantic relation spaces by projecting entities into relation-specific subspaces and learning distinct relation weight distributions. To further suppress noise from high-order message passing, we introduce a contrastive regularizer that leverages multi-relation subgraph variants to enforce consistency across augmented views. Moreover, we develop a personalized denoising encoder that dynamically refines user-item representations end-to-end, removing the need for external data generation modules. We evaluate GAT++ on extensive real-world datasets across music, literature, and food domains. GAT++ achieves up to 34.81% improvement in Recall@N over strong baselines, demonstrating its effectiveness and generalizability across diverse recommendation scenarios.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 22076