Link Property Prediction - ogbl-biokg

This code includes implementation of many KG models with OGB evaluator. It is based on this repository.

Authors: Arshjot Singh Khehra and Srihari Sridharan

The repository contains working implementation for the following models:

• TransE (same as reference repository)
• DistMult (same as reference repository)
• ComplEx (same as reference repository)
• RotatE (same as reference repository)
• ConvE (use branch 1-to-N as it contains implementation with 1-N scoring)
• NormConvKB (ConvKB with Lp normalization instead of Batch Normalization)
• TransM (RESCAL)
• HolE
• RelConv
• QuatE
• ConvFM (Our custom model)
• ConEx
• ConvQuatE (Our custom model)
• OctonionE

Results

The top scoring results (only for models implemented in addition to the reference) have been summarized below:

Model Architecture	Details	Valid MRR	Test MRR
HolE	Effective batch size = 1024, starting lr = 0.001	0.81384	0.81374
QuatE	Effective batch size = 1024, starting lr = 0.0005	0.81169	0.81104
OctonionE	Effective batch size = 1024, starting lr = 0.0005	0.81218	0.81058
NormConvKB	Effective batch size = 1024, starting lr = 0.0005	0.76682	0.76697

For training the HolE, OctonionE, and QuatE models, run:

bash examples.sh

References

• PyKEEN Library
• Reference implementation for ogbl-kg