xingye1999 / RNCE

The implementation for paper ''Learning Regularized Noise ContrastiveEstimation for Robust Network Embedding''

This repo includes the implementation of methods with RNCE and an re-implementation of RNS.

Files

We adopt the proposed regualrizer on several methods including LINE, node2vec, VERSE. The files are:

• base-rnce.cpp (for vanilla-x, vanilla-h, LINE-1st, LINE-2nd, LINE)
• node2vec-rnce.cpp (for node2vec)
• verse-rnce-sgd.cpp (for verse with SGD)
• verse-rnce-mbsgd.cpp (for verse with Mini-batch SGD)
• rns.cpp (for RNS)

Compile

sh compile.sh

RNCE

The implementation of TKDE'22 paper ''Learning Regularized Noise ContrastiveEstimation for Robust Network Embedding''.

Usage Example

./$method -input network_path -reg $distanc_func -beta 0.01 -emb-u $emb_x_path -emb-v $emb_h_path -samples 100 -negatives 5 

For full usage, please refer to the main() function in cpp code.

RNS

We re-implement the paper AAAI19 paper ''Robust Negative Sampling for Negative Embedding'' (including the embedding penalty and the adaptive negative sampler).

Usage example

./rns -input $network_path -rns 3 -emb-u $emb_x_path -emb-v $emb_h_path -samples 100 -negatives 5

Notes:

• rns=0 for without rns
• rns=1 for only embedding norm penalty
• rns=2 for only adapative negative sampler
• rns=3 for both embedding norm penalty and adapative negative sample

Experiments

Node Classification

Please refer to code

Network Reconstruction

First,

cd eval_network_reconstruction

and then,

python network_reconstruction.py --emb1 $emb_x_path --emb2 $emb_h(x)_path --net $network_path

Citation

@ARTICLE{rnce,
  author={Xiong, Hao and Yan, Junchi and Huang, Zengfeng},
  journal={IEEE Transactions on Knowledge and Data Engineering}, 
  title={Learning Regularized Noise Contrastive Estimation for Robust Network Embedding}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TKDE.2022.3148284}}