Towards Semantic Understanding of Graph Neural Network Layers Embedding with Functional Semantic Activation Mapping

Graph Neural Networks (GNNs) are now a standard tool for modelling graph structured data in applications such as molecular property prediction, drug discovery, recommender systems, and citation networks. However, despite their strong predictive performance, they still suffer from the black box problem. Most existing explainability methods focus on local-level explainability, explaining individual predictions. They highlight important nodes and edges but don′t capture how the model behaves globally across a dataset. As a result, global-level explainability remains an open challenge. In this paper, we extend our previous work on Functional Semantic Activation Mapping (FSAM) to investigate how varying the number of GNN layers affects both representation quality and predictive performance. Across several datasets, increasing depth may improve accuracy but does not necessarily enhance semantic coherence. In some cases, performance gains coincide with a decline in semantic quality, suggesting that spurious patterns may drive correct predictions for wrong reasons. FSAM layer-wise activation tracking allowed us to track neuron activations across layers, revealing that deeper layers can reduce neuron specialisation and lead to class misclassifications. Our findings demonstrate a critical trade-off that increased depth can compromise interpretability without commensurate gains in meaningful semantic learning.