Graph attention-based neural collaborative filtering for item-specific recommendation system using knowledge graph

Recently, the use of graph neural networks (GNNs) for leveraging knowledge graphs (KGs) has been on the rise due to their ability to encode both first-order and higher-order neighbor information. Most GNN-based models explicitly encode first-order information of an entity but may not effectively capture higher-order information. To address this, many existing methods overlook the impact of varying relations among neighboring nodes, leading to the integration of nodes with diverse semantics. This work propose an end-to-end recommendation model, named Item-Specific Graph Attention Network (IGAT), which jointly utilizes user-item interaction and KG information to predict user preferences. IGAT incorporates a knowledge-aware attention mechanism that assigns different weights to neighboring entities based on their relations and latent vector representations in the KG. Additionally, an item-specific attention mechanism is applied to measure the influence of the target item on the user's historical items. To mitigate biases from multi-layer propagation, IGAT utilizes contextualized representations of both users and items in the recommendation process. Extensive experiments on three benchmark datasets demonstrate the superior performance of IGAT compared to state-of-the-art KG-based recommendation models, with results showing that the proposed model outperforms the baselines. © 2024 Elsevier Ltd