TY - JOUR

T1 - Multi-feature Adaptive-fusion Enhanced graph neural Network for open-set node classification

AU - Liu, Xinxin

AU - Cao, Jie

AU - Yu, Weiren

AU - Li, Zongxi

AU - Wang, Yuyao

AU - Gu, Huanhuan

AU - Вукович, Дарко

N1 - Liu, X., Cao, J., Yu, W., Li, Z., Wang, Y., Gu, H., & Vuković, D. B. (2025). Multi-feature adaptive-fusion enhanced graph neural network for open-set node classification. Neurocomputing, 130238.

PY - 2025/8/1

Y1 - 2025/8/1

N2 - Semi-supervised open-set node classification is vital in graph learning, aiming to classify seen class nodes while identifying unseen class ones. Existing methods face limitations: they use fixed weight or simple concatenation for multi-view fusion, ignoring node-specific heterogeneity and inter-view relationships, and rely on global thresholds for novelty detection, which overlooks class-wise distribution variations. To address these issues, we introduce a Multi-feature Adaptive-fusion Enhanced graph neural Network (MAEN) for open set node classification. MAEN comprises two key components: (1) A view-aware adaptive fusion mechanism that dynamically integrates multi-view features using a mixture-of-experts-guided weight generation strategy, effectively capturing node-specific characteristics and nuanced inter-view dependencies. (2) A distribution aware rejection strategy that constructs class-adaptive decision boundaries by modeling the probability density distributions of seen classes, ensuring precise identification of novel-class instances. Experiments show that MAEN outperforms baseline methods across benchmark datasets, achieving significant improvements in both closed-set and open-set tasks.

AB - Semi-supervised open-set node classification is vital in graph learning, aiming to classify seen class nodes while identifying unseen class ones. Existing methods face limitations: they use fixed weight or simple concatenation for multi-view fusion, ignoring node-specific heterogeneity and inter-view relationships, and rely on global thresholds for novelty detection, which overlooks class-wise distribution variations. To address these issues, we introduce a Multi-feature Adaptive-fusion Enhanced graph neural Network (MAEN) for open set node classification. MAEN comprises two key components: (1) A view-aware adaptive fusion mechanism that dynamically integrates multi-view features using a mixture-of-experts-guided weight generation strategy, effectively capturing node-specific characteristics and nuanced inter-view dependencies. (2) A distribution aware rejection strategy that constructs class-adaptive decision boundaries by modeling the probability density distributions of seen classes, ensuring precise identification of novel-class instances. Experiments show that MAEN outperforms baseline methods across benchmark datasets, achieving significant improvements in both closed-set and open-set tasks.

UR - https://www.sciencedirect.com/science/article/pii/S0925231225009105

U2 - 10.1016/j.neucom.2025.130238

DO - 10.1016/j.neucom.2025.130238

M3 - Article

VL - 640

JO - Neurocomputing

JF - Neurocomputing

SN - 0925-2312

M1 - 130238

ER -