voetberg / xnugraph Goto Github PK

• Class to take a trained network, make an instance of it, and apply an arbitrary explainablity/interpretability method (passed as a string w/ param dictionary or pointer to class instance)
• Should have the ability to work with single datapoints or batches of data
• Functionality with Slurm for WC submission (plugin from pl for slurm access ?? Need to do some reading)

Should be able to extract the weighted edges from the model and make a visualization

Show a well performing example and a poorly performing example - This can also be used for subgraphs in the future when subgraph based explanations are set up.

Complexity comes from the 3 planes of the output - Need to be able to show all of them on one image without it becoming too cluttered

Impliment gnn-lrp on nugraph - work into the abstract class

reference paper - https://arxiv.org/pdf/2006.03589.pdf

• Include weights of nexus connections in some sort of histogram (?) Or any general visualization (to see the impact of nexus nodes on predictions
• Specialized masks for incorrect results (Special classes for y_hat != y predictions)
• Ensure uniform scaling for all plots

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

integrate into existing framework

Impliment XGNN on NuGraph - include visuals and local inference outputs.

Reference paper - https://dl.acm.org/doi/pdf/10.1145/3394486.3403085

Should be usable with the module from #1

Edges in the graph are not bidirectional - it just draws them twice.
All drawing programs and edge masking should mask both directions or only optimize over one direction.

Probably solvable with changing all the edges to sets but hey

When a subgraph is displayed, a ghost of the whole graph should be under the subgraph.
There should also be a uniform axis (x,y) for all subgraphs of the same graph. It'll make it easier to compare which parts of the graph compose the subgraph.

• Make a subset of the data with obvious explanations, and show that the methods pick this up

(Only one or two graphs?)

• Abstract class with visualization and inference methods to apply an interpretablity method to a single/batch of data

After discussion in the workshop - it would be advantageous to be able to monitor the message passing after each iteration

• During training -> include a static decoder that looks at the messages after each iteration (basically, an advanced loss curve).
• After training -> Stick this static decoder in after each step of linear planes, see where the most important part of the network is (similar to a relevance walk)

• Swap the gnnexplainer around to specifically pick the most important features in the network for each node/graph
  • Should produce some sort of bar graph that shows when you hover over each node
  • Additionally meta-measure (mean importance of features across all nodes in graph, mean importance over whole dataset - split by tagged events)

Explanations expect network to output a single class (or the probabilities for all classes) - so the output scheme of nugraph does not work automatically. There are a few possible solutions to this

• Rewrite the explanation method to work with multi-output
• Get an explanation for each plane of the detector (e.g. each decoder has its own explanation)

Second is way easier but the first is more rigorous and would give better results (most likely?) - it's worth exploring I'd say.