Dynamically Pruned Message Passing Networks

The code is based on our pervious naming scheme of our model. Then, we changed a lot of terms that might cause confusion when referring to components in our model in order to give a better and clearer statement in our paper. Here is the list of terminology between the previously used and the currently used:

• unconsciousness flow (previous): IGNN (now)
• consciousness flow: AGNN
• attended nodes: nodes in the attending-from horizon
• seen nodes: nodes in the attending-to horizon
• memorized nodes: visited nodes
• scanned edges: edges of neighborhood

./run.sh --dataset <Dataset>

can be one of 'FB237', 'FB237_v2', 'FB15K', 'WN18RR', 'WN18RR_v2', 'WN', 'YAGO310', 'NELL995'.

Run with test_output_attention to get data files of extracted subgraphs for each query in test. For example, if you want to get data files on the NELL995 dataset (containing several separate datasets), with max_attended_nodes=20, run:

./run.sh --dataset NELL995 --test_output_attention --max_attended_nodes 20 --test_max_attended_nodes 20

Then, you get data files in the output/NELL995_subgraph directory. Next, visualize them by:

cd code
python visualize.py --dataset NELL995

You will find image files for visualization in the visual/NELL995_subgraph directory.