Yuqicheng/KGCP-KGE-Checkpoints

KGCP: trained KGE models

45 trained knowledge graph embedding (KGE) models: DistMult, ComplEx and RESCAL on the Nations, UMLS and Kinship benchmarks, 5 seeds each. They are the models of the conformal prediction experiments of CP4KGE, a library that turns the scores of a KGE model into answer sets that contain the true answer with a chosen probability (e.g. 90%).

Each configuration was tuned per dataset with the protocol of LibKGE's authors (Ruffinelli et al., ICLR 2020), adapted to small graphs. The models were trained with LibKGE, patched for recent PyTorch and Apple Silicon.

Files

models/<dataset>-<model>/
  config.yaml           LibKGE configuration; seed<N>.pt was trained with
                        kge start config.yaml --random_seed.default <N>
  seed1.pt … seed5.pt   LibKGE packaged models: weights, configuration, entity/relation ids