Year: 2024
Type: article
Abstract: Most existing recommendation models based on knowledge graphs and contrastive learning employ random augmentation schemes to enhance the data in knowledge graphs; however, noise in knowledge graphs can lead to inaccurate recommendation results.Furthermore, most contrastive learning methods are only applied between one or two views, thereby failing to exploit the semantic information in the data fully.Therefore, this model proposes a recommendation model, MVKGCL, which integrates knowledge graphs and contrastive learning mechanisms.Firstly, it incorporates random noise into attention weights to conduct contrastive learning among different attention weights, and subsequently introduces a novel automatic masking mechanism to augment the Knowledge Graphs, performing local contrastive learning on the derived user and item embeddings.Secondly, it employs Graph Attention Network to encode the user-item-entity graph, yielding representations for users and items.Lastly, global level contrastive learning is conducted between the locally learned user and item embeddings and the node embeddings from the user-item-entity graph, uncovering comprehensive graph features and structural information.Experiments demonstrate that the model outperforms others on the Amazon-book and Yelp2018 datasets.
Authors: Zhiyue Xiong, Hankiz Yilahun, Sadiyagul Anwer, Askar Hamdulla
Institution: Xinjiang University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4406357823