I’m diving into a research project that involves building Direct Follows Graphs (DFGs) from event logs, where each node represents an activity. My goal is to use Graph Neural Networks (GNNs) to predict the next activity in a trace, and then explain the model’s decision using GNN explainability techniques.
I’ve managed to create prefix graphs from the DFGs and applied a GNN for next activity prediction. Now, I’m stuck on explaining the model’s predictions, especially since the DFGs are mostly linear with some self-loops or a few normal loops. I’m wondering if it makes sense to search for subgraphs that explain the result using perturbation-based methods or surrogate-based methods.
For instance, if I use a 3-layer GNN, wouldn’t the prediction already be fully explained by the 3-hop neighborhood? These graphs aren’t massive, so maybe I’m missing something. As a newcomer to the world of GNNs, I’d love some insight from experts in the field.
GNN explainability techniques can be powerful tools in understanding how our models work, and I’m excited to explore this further. If you have any experience with explaining GNN predictions, I’d love to hear your thoughts.