ReliK is a measure to capture reliability in Knowledge Graph Embeddings (KGE) in a local neighborhood.
A Knowledge graph is considered a list of triples in the structure (head, relation, tail)
e.g. ("Leonardo da Vinci", "painted", "Mona Lisa")
The code is in python and located in the "approach" folder. The requirements are noted in requirements.txt and can be installed with that help.
To run the code:
python experiment_controller.py -d [dataset] -e [embedding] -t [task_list] -s [size_subgraphs] -n [nmb_subgraphs] -heur [heuristic] -r [sample_size] -c [classifier_type]
The task_list schould be seperated by ,.
To run the ReliK calculations for the CodexSmall dataset with the TransE embedding would look like:
python experiment_controller.py -d CodexSmall -e TransE -t relik -s 60 -n 100 -heur binomial -r 0.1
If there is no pretrained embedding that fits the dataset and type, it will be trained before continuing with the experiment. You can also pretrain the embedding with:
python experiment_controller.py -d [dataset] -e [embedding] -st
If additional info is needed
python experiment_controller -h
provides more info, how to run experiments.
DataSplits for reproducability will be stored in a KFold folder.
The trained embeddings will be stored in a trainedEmbeddings folder.
The resulting data will be stored in a scoreData folder.
If there are any questions about this, feel free to contact: maximilian.egger[at]cs.au.dk