Skip to content

Latest commit

 

History

History
32 lines (29 loc) · 2.85 KB

README.md

File metadata and controls

32 lines (29 loc) · 2.85 KB
Raw

Model-agnostic Graph Explainability from Scratch

This repository holds a graph explainability solution, which extends the work (GraphMask Explainer) to heterogeneous as well as homogeneous Graphs, making this functionality model-agnostic. Moreover, this implementation provides both node feature-level and edge-level attributes mask (explanation subgraph), which is a binary-valued vector. All 0 values of this mask vector represent those features (and edges) of the graph that do not affect their corresponding predictions, whereas features (and edges) associated with 1 values consider to be a lot effective in influencing their associated predictions outputted by the original Graph Neural Network (GNN) model.

Requirements

  • PyTorch Geometric
  • PyTorch
  • numpy
  • scikit-learn
  • tqdm

Usage

Data

This implementation of GraphMask Explainer demonstrates explainability examples for GCN, GAT, and RGCN layer-types on Node Classification (NC), Graph Classification (GC), and Link Prediction (LP) tasks.

Layer Type Task Dataset
GCN NC Cora
GCN GC Enzymes
GAT NC Cora
GAT GC Enzymes
GAT LP Cora
RGCN NC AIFB
RGCN GC Enzymes

Training and Testing

  • To see the model-agnostic explainability layer’s implementation, check graphmask_explainer.py.
  • To train the GraphMask Explainer and generate explanations for any of the aforementioned tasks, run graphmask_explainer_example.py.
  • All hyperparameters’ settings can be tweaked (based on requirements) by altering their corresponding values provided in both graphmask_explainer.py and graphmask_explainer_example.py files.

Results

NC Task Explanation Subgraph (AIFB Dataset) GC Task Explanation Subgraph (Enzymes Dataset)
alt text alt text

These figures show output subgraph in which all irrelevant edges (having 0 values in the binary-valued mask) have been colored grey, whereas all relevant edges (having 1 values in the generated binary-valued mask) have been illustrated in black color. Note that, for NC task, the output subgraph contains only those nodes that lie within the 3-hop neighborhood of the parent node with index 20 and have the same relation type as the parent node has, on the other hand, for GC task, the output subgraph demonstrates explanations of the graph with index 10 present in Enzymes dataset.