View in EDS

Self-supervised category-enhanced graph neural networks for recommendation.

Saved in:
Bibliographic Details
Title: Self-supervised category-enhanced graph neural networks for recommendation.
Authors: Yang, Funing1 (AUTHOR) yfn@jlu.edu.cn, Du, Haihui1 (AUTHOR) duhh22@mails.jlu.edu.cn, Zhang, Xingliang2 (AUTHOR) zhangxingliang@jl.chinamobile.com, Yang, Yongjian1 (AUTHOR) yyj@jlu.edu.cn, Wang, Ying1 (AUTHOR) wangying2010@jlu.edu.cn
Source: Knowledge-Based Systems. Feb2025, Vol. 311, pN.PAG-N.PAG. 1p.
Subjects: Graph neural networks, Recommender systems, Knowledge graphs
Abstract: The emergence of Graph Neural Networks (GNNs) has substantially advanced recommendation systems based on collaborative filtering. Despite their effectiveness, these networks are often susceptible to noise and sparsity problems from datasets. Thus, contrastive learning and knowledge graphs have been introduced as solutions to recommendation systems. However, both of these techniques have been incorporated into recommendation models as auxiliary tasks, neglecting the enhancement of node representations for the overall task. Therefore, we propose a novel contrastive learning method called C ategory- E nriched C ontrastive L earning (CECL) to obtain high-quality user–item interactions. Unlike previous studies, CECL does not introduce categories into the recommendation system as auxiliary supervised signals but uses them as direct supervised signals during model training. This directly supervised signal can enable the model to learn and emphasise category-specific patterns and relationships more efficiently, resulting in more accurate and stable user and item embeddings. By showing optimised category alignment, CECL improves representation quality. Specifically, we first construct initial embeddings for users and items based on the categories of user interactions and those to which items belong. Then, these embeddings are input into the recommendation model. The primary recommendation task is conducted in the supervised model. In the self-supervised model, contrastive pairs are constructed using edge dropout, and the consistency between different views of the same node is maximised to explore stable category preferences. Finally, multi-task training is applied to jointly optimise the recommendation and contrastive learning tasks. Comprehensive experiments conducted on eight real-world datasets validate the effectiveness of the proposed CECL. Compared with existing graph-enhanced methods, CECL achieves an average improvement of 15.17% in Recall@20 and 9.64% in NDCG@20 metrics. Our implementations are available at https://github.com/shark-art/Code. [ABSTRACT FROM AUTHOR]
Copyright of Knowledge-Based Systems is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:The emergence of Graph Neural Networks (GNNs) has substantially advanced recommendation systems based on collaborative filtering. Despite their effectiveness, these networks are often susceptible to noise and sparsity problems from datasets. Thus, contrastive learning and knowledge graphs have been introduced as solutions to recommendation systems. However, both of these techniques have been incorporated into recommendation models as auxiliary tasks, neglecting the enhancement of node representations for the overall task. Therefore, we propose a novel contrastive learning method called C ategory- E nriched C ontrastive L earning (CECL) to obtain high-quality user–item interactions. Unlike previous studies, CECL does not introduce categories into the recommendation system as auxiliary supervised signals but uses them as direct supervised signals during model training. This directly supervised signal can enable the model to learn and emphasise category-specific patterns and relationships more efficiently, resulting in more accurate and stable user and item embeddings. By showing optimised category alignment, CECL improves representation quality. Specifically, we first construct initial embeddings for users and items based on the categories of user interactions and those to which items belong. Then, these embeddings are input into the recommendation model. The primary recommendation task is conducted in the supervised model. In the self-supervised model, contrastive pairs are constructed using edge dropout, and the consistency between different views of the same node is maximised to explore stable category preferences. Finally, multi-task training is applied to jointly optimise the recommendation and contrastive learning tasks. Comprehensive experiments conducted on eight real-world datasets validate the effectiveness of the proposed CECL. Compared with existing graph-enhanced methods, CECL achieves an average improvement of 15.17% in Recall@20 and 9.64% in NDCG@20 metrics. Our implementations are available at https://github.com/shark-art/Code. [ABSTRACT FROM AUTHOR]
ISSN:09507051
DOI:10.1016/j.knosys.2025.113109