Graph Neural Networks Explainability in Social Networks: Evaluation of Methods and Feature Relevance

Identifier:  TFM:2410

Handle: https://hdl.handle.net/20.500.11797/TFM2410

Authors:  Garcia Rodriguez, Uxue

Abstract:
This work analyses the predictive capabilities of a GNN on a social network and its later explanatory abilities. Network nodes were labeled based on the augmented features of their first-level neighbors. GraphSAGE and GCN were trained using only the augmented features and on invented and augmented features. Then, explainability was assessed using GNNExplainer, PGExplainer, and GraphMask, and compared against ground-truth logic using fidelity, unfaithfulness, and characterization metrics. GNNExplainer proved most effective, closely aligning with the labeling rules. The results highlight both the strengths and limitations of these explainers in capturing meaningful structural patterns in the graph data.