Find connected nodes & edges from specific node index

[v-ngangarapu]

Hi,
I have trained a GNN model with multiple graphs that can predict node and edges. Using the trained model, I like to find in a graph, how the nodes and edges are connected from a specific node. I see some examples with CaptumExplainer that is able find the connected nodes and edges and ploting it. How I can get list of connected edges & nodes so that I can use it for further analysis?

In the screen shot provided, the explainer able to identify 476, 10 & 2545 as connected, remaining as masked and one node is isolated. I want list of connected nodes and edges only and ignore remaining.

Any suggestions here.

[wsad1]

The explainer returns and edge_mask and node_mask you can use this to obtain a list of edges/nodes that explain the prediction.
See here for an example.

[v-ngangarapu]

Thanks for directing me to the tutorials. It was helpful. I understand based on intesity of weights in mask, the plotting of edges are getting varied. But I see similar weights for different node index within the graph. It's causing issue to filter the connected nodes.

code I am using:

explainer = Explainer(
model=model,
algorithm=CaptumExplainer('IntegratedGradients'),
explanation_type='model',
node_mask_type='attributes',
edge_mask_type='object',
model_config=dict(
mode='multiclass_classification',
task_level='node',
return_type='log_probs',
),
)

node_index=1
explanation = explainer(dataset.data.x, dataset.data.edge_index.T, index=node_index)

got this plot for node_index : 8
Here, I want to select nodes - 7, 4

got this plot for node_index : 1
Here, I want to select nodes - 8, 7, 4

[rusty1s]

Sorry for late reply. PyG is not a graph analytics library, so we don't provide such functionality. I would suggest you can convert your Explanation to networkx and find meaningful subgraph/connected components from there. Would that work for you?

[v-ngangarapu]

Sorry, I got busy with other priorities and not able to reply to this.

For my research, want to solve this problem with neural network instead of networkx. I believe extending the PyG with this feature may open the door for several other use cases. I got one suggestion to apply probabilities on results, do you think any of this kind of logic will help here?

[rusty1s]

What do you mean by applying probabilities on results?

[v-ngangarapu]

It's same like measuring the weight of results. Similar to whatever explainer does. I was trying to get something with this approach. This was not helping me.