The Structural EMBedding graph library (SEMB)

Authors: GEMS Lab Team @ University of Michigan (Mark Jin, Ruowang Zhang, Mark Heimann)

This SEMB library allows fast onboarding to get and evaluate structural node embeddings. With the unified API interface and the modular codebase, SEMB library enables easy intergration of 3rd-party methods and datasets.

The library itself has already included a set of popular methods and datasets ready for immediate use.

• Built-in methods: node2vec, struc2vec, GraphWave, xNetMF, role2vec, DRNE, MultiLENS, RiWalk, SEGK, (more methods to add in the near future)

• Built-in datasets:

  Dataset	# Nodes	# Edges
  BlogCatalog	10,312	333,983
  Facebook	4,039	88,234
  ICEWS	1,255	1,414
  PPI	56,944	818,786
  BR air-traffic	131	1,038
  EU air-traffic	399	5,995
  US air-traffic	1,190	13,599
  DD6	4,152	20,640
  Synthetic Datasets

The library requires *Python 3.6.2 for best usage. In Python 3.8, the Tensorflow 1.14.0 used in DRNE might not be successfully installed.

Installation and Usage

Make sure you are using Python 3.6+ for all below!

1. First, creat a virtual environment and activate the virtual environment using conda

   conda create -n "<VENV_NAME>" python=3.6.2 ipython
   conda activate <VENV_NAME>

2. Change directory to the StrucEmbeddingLibrary and install the dependencies

   (<VENV_NAME>) cd StrucEmbeddingLibrary
   (<VENV_NAME>) python3 -m pip install -r requirements.txt --no-cache-dir

3. Install the SEMB package

   (<VENV_NAME>) cd StrucEmbeddingLibrary
   (<VENV_NAME>) python3 setup.py install

   After installation, we highly recommend you go through our Tutorial to see how SEMB library works.

4. To enable using the jupyter notebook, do the following,

   (<VENV_NAME>) python3 -m pip install ipykernel --no-cache-dir
   (<VENV_NAME>) python3 -m ipykernel install --name=<VENV_NAME>
   (<VENV_NAME>) jupyter notebook

   Choose <VENV_NAME> at the top right corner of the page when creating a new jupyter notebook / running the tutorial notebook.

Extending SEMB

First make sure the semb library is installed.

Developing 3rd party Dataset extension

Currently, SEMB only supports embedding and evaluation on undirected and unweighted graphs.

• Create a Python 3.6+ package with a name in form at semb/datasets/[$YOUR_CHOSEN_DATASET_ID]
• Within the package root directory, make sure __init__.py is present
• Create a dataset.py and make a Dataset class that inherits from from semb.datasets import BaseDataset and implement the required methods. See semb/datasets/airports/dataset.py for more details.
  • To use the built-in load_dataset()method, we accept the graph edgelist with the following format
    • <Node1_id (int)> <Blank> <Node2_id (int)> <\n>
    • Otherwise, you can overload and implement your own load_dataset() function. Please make sure that the returned graph is of networkx.classes.graph.Graph datatype.
  • If the dataset is accompanied by the label file, to use the built-in load_label() function, we accept the label file with the following format
    • <Node_id (int)> <delimeter> <Node_label (int)>
    • Otherwise, you can overload and implement your own load_label() function. Please make sure that the returned type is python built-in dict() with the key as <Node_id (int)> and value as <Node_label (int)>
• Install the package via setup.py or pip.
• Now the dataset is loadable by the main client program that uses semb!

Developing 3rd party Method extension

• Create a Python 3.6+ package with a name in form of semb/methods/[$YOUR_CHOSEN_METHOD_ID]
• Within the package root directory, make sure __init__.py is present
• Create a method.py and make a Method class that inherits from from semb.methods import BaseMethod and implement the required methods. See semb/methods/node2vec/method.py for more details.
  • Please make sure that your implemented method accepts networkx.classes.graph.Graph as input.
  • Please make sure that when train() is called, the self.embeddings should be a Python built-in dict() with key as <Node_id (int)> and value(embedding) as <List (float)>.
• Install the package via setup.py or pip.
• Now the method is load-able by the main client program that uses semb!

Note

For both dataset and method extensions, make sure the get_id() to be overridden and returns the same id as your chosen id in your package name.

Contact

If you encounter any question using our SEMB library, feel free to raise an issue or send an email to kinmark@umich.edu. Go Blue!