JWalk: Representational learning on graphs¶
jwalk¶
jwalk performs random walks on a graph and learns representations for nodes using Word2Vec. It also has options to train existing models online and specify weights.
Install¶
pip install -U jwalk
Build¶
make build
Usage¶
jwalk -i tests/data/karate.edgelist -o karate.emb --delimiter=' '
To see the full list of options:
jwalk --help
Prompt parameters:
debug: drop a debugger if an exception is raised
delimiter: delimiter for input file
embedding-size: dimension of word2vec embedding (default=200)
has-header: boolean if csv has header row
help (-h): argparse help
input (-i): file input (edgelist of 2/3 cols or adjacency matrix)
log-level (-l) logging level (default=INFO)
model (-m): use a pre-existing model
num-walks (-n): number of of random walks per graph (default=1)
output (-o): file output
stats: boolean to calculate walk statistics [requires pandas]
undirected: make graph undirected
walk-length: length of random walks (default=10)
window-size: word2vec window size (default=5)
workers: number of workers (default=multiprocessing.cpu_count)
Input File¶
The input file can be of the following formats:
- Edgelist: CSV with 2 or 3 columns denoting the source, target and (optional) weight. There are CLI options to specify the delimiter and whether the file has a header (default=False). The CSV file is loaded using numpy if pandas is not installed. We strongly recommend using pandas to load the CSV as it’s a lot faster.
- Graph: If the file has an extension that is ”.npz”, jwalk will assume that it is a SciPy CSR matrix. Included must be keys of data, indices, indptr, shape and labels (default=None) where labels are the node labels. For an example, see tests/data/karate.npz.
Blog¶
Read more about jwalk in our blog post here: https://www.jwplayer.com/blog/deepwalk-recommendations/
License¶
Apache License 2.0
References¶
- [paper]: arXiv:1403.6652 [cs.SI] “DeepWalk: Online Learning of Social Representations”
- [paper]: arXiv:1607.00653 [cs.SI] “node2vec: Scalable Feature Learning for Networks”