metapath2vec

metapath2vec: Scalable Representation Learning for Heterogeneous Networks

• Paper link: https://ericdongyx.github.io/papers/KDD17-dong-chawla-swami-metapath2vec.pdf

• Author's code repo (in Tensorflow):

  https://github.com/apple2373/metapath2vec.

Dataset Statics

Dataset	# Nodes	# Node Types	# Edges	# Edge Types	Target	# Classes
AMiner	4,891,819	3	25,036,020	4	author	8
DBLP	26,128	4	239,566	6	author	4
IMDB	11,616	3	34,212	4	movie	3

Results

TL_BACKEND="torch" python metapath2vec_trainer_aminer.py --lr 0.1 --embedding_dim 16 --walk_length 60 --window_size 3 --num_walks 600 --n_epoch 5 --num_negative_samples 6 --batch_size 128 --train_ratio 0.5 --dataset aminer
TL_BACKEND="torch" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset imdb
TL_BACKEND="torch" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset dblp
TL_BACKEND="tensorflow" python metapath2vec_trainer_aminer.py --lr 0.1 --embedding_dim 16 --walk_length 60 --window_size 3 --num_walks 500 --n_epoch 5 --num_negative_samples 6 --batch_size 128 --train_ratio 0.5 --dataset aminer
TL_BACKEND="tensorflow" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset imdb
TL_BACKEND="tensorflow" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset dblp
TL_BACKEND="paddle" python metapath2vec_trainer_aminer.py --lr 0.1 --embedding_dim 16 --walk_length 60 --window_size 3 --num_walks 600 --n_epoch 5 --num_negative_samples 6 --batch_size 128 --train_ratio 0.5 --dataset aminer
TL_BACKEND="paddle" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset imdb
TL_BACKEND="paddle" python metapath2vec_trainer_imdb_dblp.py --lr 0.01 --embedding_dim 16 --walk_length 50 --window_size 7 --num_walks 5 --n_epoch 50 --num_negative_samples 5 --batch_size 128 --dataset dblp

Dataset	HAN(report)	Our(th)	Our(tf)	Our(pd)
AMiner	84.27(pyg)	84.47(±0.57)	83.54(±1.15)	84.05(±1.43)
IMDB	45.65	51.80(±0.43)	51.54(±1.23)	51.24(±1.42)
DBLP	91.53	91.76(±0.34)	91.42(±0.22)	91.56(±1.56)

Note: The hyperparameter settings for the results in the AMiner dataset is the same as in GammaGL.