基于多视图邻居对比学习的节点分类方法

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）01-017-0145-08

doi：10.19734/j. issn.1001-3695.2026.05.0170

Multi-view neighbor contrastive learning for node classification

Liu Junlong，Dong Jizhou，Wang Yidan† （HebeiKeyLboatofcadmpualteligcelofaicsftoce，beUit Baoding Hebei 071002，China）

Abstract：The widespreadapplicationof graphcontrastivelearning（GCL）methods innodeclasification has effectivelyalleviatedtherelianceonlabelinformation.However，existingGCLmethodstend toaggregatealargeamountof informationfrom dissimilar nodes inheterophilic gaphs.Moreover，apotentialconflictarisesbetweentheneighboraggregationprocessof graph neural networksandtheoptimizationobjectivesofcontrastivelossfunctions.Tomitigate these isues，thispaperconducteda theoreticalaalysis toinvestigatethecausesofperformancedegradationinheterophilicgraphsandthesourceofconflicts within graph contrastivelearming.Basedonthis analysis，it proposed MVNCL，anovel node clasification framework.Specificaly， MVNC introducedastructure-reconstruction-basedaugmentationstrategythatincorporatednodesimilarityandclassuncertaintytomore effectivelyidentifyhard negative samples.This approach generated augmented views in which connectednodes were morelikelytobelong tothesameclass，therebypromoting efectivefeatureaggregation.Inaddition，MVNCLdesigneda neighborcontrastivelossfunctionthatcomparednoderepresentationsacrosstheoriginalandaugmentedviews，aswellaswith non-structuralviews.Thisdesignreduced theconflictbetween featureaggregationandthecontrastivelearningobjective. Extensiveexperimentsonfive benchmark datasets demonstratethatMVNCLconsistentlyoutperforms existing methodsonboth homophilicand heterophilic graphs，offering an effective solution for node classfication acrossdiverse graph structures.

ey words：graph contrastive learning；node clasification；graph neural network；neighbor contrastive loss

0 引言

随着大量图结构数据的应用，节点分类任务已经得到了迅速发展，并在推荐系统[1]、社交网络分析[2]、电商营销[3]等多个现实场景中展现出广泛的应用价值。（剩余18559字）