Graph Neural Networks (GNNs) have established themselves as powerful tools for graph-structured data. However, when feature separability among nodes is low, conventional neighborhood aggregation strategies often result in performance degradation due to over-smoothing and noisy information propagation. In this work, we introduce a novel GNN framework that refines the aggregation process by integrating feature similarity and neighborhood entropy into node message passing. Unlike standard models that uniformly aggregate neighbor information, this new model dynamically adjusts neighbor influence, prioritizing nodes with high similarity and low entropy. We evaluate the model on synthetic graphs generated using the Stochastic Block Model (SBM), varying homophily and class imbalance conditions to simulate challenging classification scenarios. Our experiments demonstrate that our method consistently outperforms the GraphSAGE model across Accuracy, Balanced Accuracy, F1 Score, and Matthews Correlation Coefficient, particularly in graphs with weak class signal and high structural noise. Statistical significance tests confirm the robustness of the observed improvements. Furthermore, the optimization of aggregation weights provides information about the model’s adaptation to different graph structures, enhancing interpretability. These findings suggest that integrating structural information selectively during aggregation can significantly improve GNN performance in complex graph environments, offering a new course toward more explainable graph-based models.

Improving Node Classification for Graphs with Weak Feature Signals: A Similarity-Entropy Aggregation Approach

Brian Daniel Bernhardt;Chiara Marciano;Mario Rosario Guarracino
2025-01-01

Abstract

Graph Neural Networks (GNNs) have established themselves as powerful tools for graph-structured data. However, when feature separability among nodes is low, conventional neighborhood aggregation strategies often result in performance degradation due to over-smoothing and noisy information propagation. In this work, we introduce a novel GNN framework that refines the aggregation process by integrating feature similarity and neighborhood entropy into node message passing. Unlike standard models that uniformly aggregate neighbor information, this new model dynamically adjusts neighbor influence, prioritizing nodes with high similarity and low entropy. We evaluate the model on synthetic graphs generated using the Stochastic Block Model (SBM), varying homophily and class imbalance conditions to simulate challenging classification scenarios. Our experiments demonstrate that our method consistently outperforms the GraphSAGE model across Accuracy, Balanced Accuracy, F1 Score, and Matthews Correlation Coefficient, particularly in graphs with weak class signal and high structural noise. Statistical significance tests confirm the robustness of the observed improvements. Furthermore, the optimization of aggregation weights provides information about the model’s adaptation to different graph structures, enhancing interpretability. These findings suggest that integrating structural information selectively during aggregation can significantly improve GNN performance in complex graph environments, offering a new course toward more explainable graph-based models.
File in questo prodotto:
File Dimensione Formato  
Improving Node Classification for Graphs withWeak Feature Signals.pdf

accesso aperto

Descrizione: Versione editoriale
Tipologia: Versione Editoriale (PDF)
Licenza: Creative commons
Dimensione 388.39 kB
Formato Adobe PDF
388.39 kB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11580/128423
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
social impact