NetSMF: Large-Scale Network Embedding as Sparse Matrix Factorization [arxiv]

Please cite our paper if you use this code in your own work:

@inproceedings{qiu2019netsmf,
 author = {Qiu, Jiezhong and Dong, Yuxiao and Ma, Hao and Li, Jian and Wang, Chi and Wang, Kuansan and Tang, Jie},
 title = {NetSMF: Large-Scale Network Embedding As Sparse Matrix Factorization},
 booktitle = {The World Wide Web Conference},
 series = {WWW '19},
 year = {2019},
 publisher = {ACM}
} 

sudo apt-get install cmake
sudo apt-get install libgflags-dev
sudo apt-get install liblog4cxx-dev
sudo apt-get install libomp-dev
sudo apt-get install libeigen3-dev
https://github.com/xptree/NetSMF.git
cd NetSMF
mkdir build
./configure
cd build
make

The dependence versions that the code is tested:

Note: Using eigen3 3.2.5 may cause problems. Please do update you eigen3 to 3.3 or above.

Support undirected networks with edgelist format.

For unweighted networks, each edge should appear twice a b and b a.

For weighted networks, each edge should appear twice a b w and b a w.

You may want to use example/mat2edge.py to translate mat to edgelist.

.mat files can be downloaded here:

• BlogCatalog Source Preprocessed
• Protein-Protein Interaction Source Preprocessed
• Flickr
• YouTube

For unweighted networks, see example/blog.sh for an example.

blog.sh takes three arguments, the first one indicates the input edgelist file, the second one indicates the output file, the third one indicating the origin .mat file containing network and labels.

For exmaple, runing ./blog.sh blogcatalog.edgelist blogcatalog.netsmf blogcatalog.mat will

• check if blogcatalog.edgelist is a valid file. If not, it calls mat2edge.py to translate mat file blogcatalog.mat to edgelist blogcatalog.edgelist.
• call NetSMF algorithm, and store the 128-dim embedding at blogcatalog.netsmf_128.npy.
• call predict.py to evaluate NetSMF at the label classification task.

You can use -weight to switch to weighted networks and use -noweight to switch to unweighted network (default unweighted).

We propose to use truncated logarithm in our WWW'19 paper.

In the code, we provide a new option log1p, i.e., log(1+x). You can use -log1p to turn it on and -nolog1p to turn it off (default off). Empirically speaking, log1p sometimes achieves better performance, for example in wiki dataset.

The implementation of randomized singular value decomposition is by redsvd and HPCA.