<p>Social recommendation, aiming to integrate social relations and user–item interactions information to enhance the recommendation performance, has received increasing attention in academia and industry. Recently, graph neural networks (GNNs)-based methods for social recommendation have achieved remarkable performance. However, most of them often overlook two key problems related to social relations. Firstly, social relations often contain lots of noises that users with friendships may not necessarily have similar preferences. Secondly, social relations are often very sparse that most users have only a few friends. These two problems prevent GNNs from obtaining accurate and sufficient social supervised signals, resulting in the suboptimal performance. In view of this, we propose a <Emphasis Type="BoldUnderline">S</Emphasis>elf-<Emphasis Type="BoldUnderline">S</Emphasis>upervised <Emphasis Type="BoldUnderline">S</Emphasis>ocial <Emphasis Type="BoldUnderline">R</Emphasis>elations <Emphasis Type="BoldUnderline">R</Emphasis>efinement (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\text {S}^3\text {R}^2$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mtext>S</mtext> <mn>3</mn> </msup> <msup> <mtext>R</mtext> <mn>2</mn> </msup> </mrow> </math></EquationSource> </InlineEquation>) method for social recommendation. Specifically, this method designs an adaptive denoising network, which can automatically adjust the relations weights under the supervision of collaborative signals and highlight preference-relevant social relations. To alleviate the sparsity problem, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\text {S}^3\text {R}^2$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mtext>S</mtext> <mn>3</mn> </msup> <msup> <mtext>R</mtext> <mn>2</mn> </msup> </mrow> </math></EquationSource> </InlineEquation> constructs noisy view and hypergraph view to add more auxiliary self-supervised signals and simultaneously utilizes multi-view contrastive learning framework to fully train the recommendation model. Extensive experiments conducted on three benchmark datasets demonstrate the superiority of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\text {S}^3\text {R}^2$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mtext>S</mtext> <mn>3</mn> </msup> <msup> <mtext>R</mtext> <mn>2</mn> </msup> </mrow> </math></EquationSource> </InlineEquation> over state-of-the-art methods.</p>

Enhancing social recommendation via self-supervised social relations refinement

期刊论文
发表日期：

2026年7月13日

查看全文

Chaobo He,
Xinran Chen,
Feiyu Peng,
Peng Mei,
Huijuan Hu,
Quanlong Guan

摘要

Social recommendation, aiming to integrate social relations and user–item interactions information to enhance the recommendation performance, has received increasing attention in academia and industry. Recently, graph neural networks (GNNs)-based methods for social recommendation have achieved remarkable performance. However, most of them often overlook two key problems related to social relations. Firstly, social relations often contain lots of noises that users with friendships may not necessarily have similar preferences. Secondly, social relations are often very sparse that most users have only a few friends. These two problems prevent GNNs from obtaining accurate and sufficient social supervised signals, resulting in the suboptimal performance. In view of this, we propose a Self-Supervised Social Relations Refinement ( $\text {S}^3\text {R}^2$ $S^{3} R^{2}$ ) method for social recommendation. Specifically, this method designs an adaptive denoising network, which can automatically adjust the relations weights under the supervision of collaborative signals and highlight preference-relevant social relations. To alleviate the sparsity problem, $\text {S}^3\text {R}^2$ $S^{3} R^{2}$ constructs noisy view and hypergraph view to add more auxiliary self-supervised signals and simultaneously utilizes multi-view contrastive learning framework to fully train the recommendation model. Extensive experiments conducted on three benchmark datasets demonstrate the superiority of $\text {S}^3\text {R}^2$ $S^{3} R^{2}$ over state-of-the-art methods.