IS-DEA Code Repository

We provide our proposal implementation and baseline implementations here.

• Our implementation is under src.
• Our baseline reproduction is under clone/*. Baselines are cloned from repos given in original paper with essential modifications to run with our data. All modifications are marked by comment # MODIFY and # \\:.

To execute our code, a virtual environment created by CONDA is required. Then execute environment.sh to install all dependencies. Execute pip install -e . to install our implementation as importable Python module etexood under debugging mode.

To execute baseline codes, please follow the README in each baseline repos to create independent virtual environments.

Execute

All example executions (run.sh) will run the least amount epoch (mostly 1). To reproduce results, please follow configuration in baseline papers and our paper. We use 50 epochs for all experiments.

In run.sh, we provide the simplest example for synthetic FD-1 and FD-2 tasks. To run other real-world inductive knowledge graph completion datasets, please refer to share-generate.sh, share-fit.sh and share-transform.sh. For example, to run an experiment with ISDEA using mean DSS aggregation on FB237 v1, please set following variables in those three scripts as

task=FB2371
model=dssgnn
aggr=mean
ablate=both

To execute any baseline, go to corresponding directory under clone. In the directory, run link.sh first to link datasets into baseline directory, then run example execution run.sh. If you want to run on a new datasets, please change task variable in the example scripts of baselines.