LMH-KGR : Learning Multi-hop Path for Multi-hop Knowledge Graph Reasoning

• The paper is under WWW review.

Encoder

• Pata sampling and path encoding.
• The datasets consist of umls, kinship, WN18RR, NELL-995, and FB15K-237.
• The number of paths sampled and the ratio include 2, 5, 10 and 1,2,3, respectively.
• For the convolution-based encoder on the umls dataset with the number of valid path 2 and ratio 3,

python encoder_conv_2.py --dataset umls  --number_multihop_valid_paths 2 --ratio_valid_invalid 3  --cuda 0

Decoder

• Train the multi-hop reasoning decoder.

./experiment-rs.sh configs/<dataset>-rs.sh --train <gpu-ID>

Acknowledgement

We refer to the code of MultihopKg here. Thanks for their contributions.