GNNExplainer Inference

[voetberg]

reference paper - https://arxiv.org/pdf/1903.03894.pdf

integrate into existing framework

[voetberg]

Pytorch geometric has this already implemented https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.explain.algorithm.GNNExplainer.html

[voetberg]

Needs to be tested on data, visuals iterated on

[voetberg]

Example Visuals

[voetberg]

Update to include the nexus node connections in the training -

[voetberg]

Updated with filtering using topk hard thresholding

[voetberg]

Applied this method to the actual network

Negative Fidelity metric is Miserable - so there are multiple courses of action

• Change the way thresholding is done - instead of taking the top 1/3 importances, take all edges above a certain quartile, instead of using straight topk.
• Train a different explainer for each class - make a mask on edges for each class and train those individually

[voetberg]

Updating the method to train for the top quartile instead of a topk. - Shows an improvement in characterization

[voetberg]

Modified to create a secondary method that only works on each class (optimized results for each class separately). Results are as follows:

MIP

HIP

Michel

diffuse

[voetberg]

Current updates - Targeting a class in the loss function for the mask optimizer, but not limiting the edges the filter can act on.

Showing fairly good positive and negative fidelity. Doing a full training run on the positive and negative examples. Loss curve and standard test sample below.

[voetberg]

Further investigation

• evenly split events vs unevenly split events (How important the nexus plane is, do events with a heavy split of hints between different planes have a stronger need for )
• correlation (scatter) of importance of edges vs length of edges (position of edge euclidean distance from node to node vs importance)
• Full histogram of edges - no cuts