ad93 / anrl

ANRL: Attributed Network Representation Learning via Deep Neural Networks(IJCAI-2018)

ANRL

ANRL: Attributed Network Representation Learning via Deep Neural Networks (IJCAI-18)

This is a Tensorflow implementation of the ANRL algorithm, which learns a low-dimensional representations for each node in a network. Specifically, ANRL consists of two modules, i.e., neighbor enhancement autoencoder and attribute-aware skip-gram model, to jointly capture the node attribute proximity and network topology proximity.

Requirements

• python2.7 or python3.6
• tensorflow
• networkx
• numpy
• scipy
• scikit-learn

All required packages are defined in requirements.txt. To install all requirement, just use the following commands:

pip install -r requirements.txt

Basic Usage

Input Data

For node classification, each dataset contains 3 files: edgelist, features and labels.

1. citeseer.edgelist: each line contains two connected nodes.
node_1 node_2 (weight)
node_2 node_3 (weight)
...

2. citeseer.feature: this file has n+1 lines.
The first line has the following format:
node_number feature_dimension
The next n lines are as follows: (each node per line ordered by node id)
(for node_1) feature_1 feature_2 ... feature_n
(for node_2) feature_1 feature_2 ... feature_n
...

3. citeseer.label: each line represents a node and its class label.
node_1 label_1
node_2 label_2
...

For link prediction, each dataset contains 3 files: training edgelist, features and test edgelist.

1. xxx_train.edgelist: each line contains two connected nodes.
node_1 node_2 (weight)
node_2 node_3 (weight)
...

2. xxx.feature: this file has n+1 lines.
The first line has the following format:
node_number feature_dimension
The next n lines are as follows: (each node per line ordered by node id)
(for node_1) feature_1 feature_2 ... feature_n
(for node_2) feature_1 feature_2 ... feature_n
...

3. xxx_test.edgelist: each line contains two connected nodes.
node_1 node_2 1 (positive sample)
node_2 node_3 0 (negative sample)
...

Output Data

The output file has n+1 lines as the input feature files. The first line has the following format:

node_number embedding_dimension

The next n lines are as follows: node_id dim_1, dim_2, ... dim_d

Run

To run ANRL, just execute the following command for node classification task:

python main.py

Note: As for simulating random walks, we directly use the code provided in node2vec, which levearges alias sampling to faciliate the procedure.

Citing

If you find ANRL useful for your research, please consider citing the following paper:

@inproceedings{ijcai2018-438,
  title     = {ANRL: Attributed Network Representation Learning via Deep Neural Networks},
  author    = {Zhen Zhang and Hongxia Yang and Jiajun Bu and Sheng Zhou and Pinggang Yu and Jianwei Zhang and Martin Ester and Can Wang},
  booktitle = {Proceedings of the Twenty-Seventh International Joint Conference on
               Artificial Intelligence, {IJCAI-18}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {3155--3161},
  year      = {2018},
  month     = {7},
  doi       = {10.24963/ijcai.2018/438},
  url       = {https://doi.org/10.24963/ijcai.2018/438},
}