HNIP

This repository provides a reference implementation of HNIP proposed in "Temporal Network Embedding with High-Order Nonlinear Information", Zhenyu Qiu, Wenbin Hu, Jia Wu, Weiwei Liu, Bo Du and Xiaohua Jia, AAAI 2020

The HNIP algorithm learns a representations for nodes in a temporal graph. Please check the paper for more details.

Basic Usage

$ python main.py -c config/xx.ini

Requirements

The implementation of HNIP is tested under Python 3.7, with the following packages installed:

• tensorflow-gpu==1.14.0
• numpy
• scipy

Input

Your input graph data should be a txt file and be under GraphData folder

Input format

The txt file should be adjlist with time stamp. In particular, the i-th line contains information about the i-th node and has the following structure: $$A:n_1,w_1,t_1;n_2,w_2,t_2;...;n_k,w_k,t_k$$

where $A$ is the node ID, $n_1,...,n_k$ are the nodes adjacent to this node, $$w_1,...,w_k$$ are the weights of the links, and $$t_1,...,t_k$$ are the time stamps of the links. Note that the nodes are numbered starting from 0. Let $t_{max}$ be the time of the last link, $t_min$ be the time of the first link, and $t_c$ be the actual time of the link $e_k$, then the $t_k = \frac{t_c - t_{min}}{t_{max}-t_{min}}$.

txt file

0:1,1.0,0.34;2,1.0,0.33;
1:0,1.0,0.34;3,1.0,0.34;
...

Output

The output is the learned representation of the input network, all lines are node ID and d dimensional representation:

0 0.0009352565 9.563565e-05 0.0013471842 ...
1 0.0009587407 9.6946955e-05 0.0013585389 ...
...

Baselines

In our paper, we used the following methods for comparision:

• DeepWalk 'Deepwalk:online learning of social representations' source
• Node2vec ' node2vec: Scalable feature learning for networks' source
• SDNE 'Structural deep network embedding' source
• CTDNE 'Dynamic network embeddings: From random walks to temporal random walks'
• NetWalk 'Netwalk: A flexible deep embedding approach for anomaly detection in dynamic networks' source

Note that the CTDNE is not open-sourced, and we have implemented it based on the published paper. You can find the implementation in the utils folder.

Citing

If you find HNIP useful in your research, we ask that you cite the following paper:

@inproceedings{DBLP:conf/aaai/QiuH00DJ20,
  author    = {Zhenyu Qiu and Wenbin Hu and Jia Wu and Weiwei Liu and Bo Du and Xiaohua Jia},
  title     = {Temporal Network Embedding with High-Order Nonlinear Information},
  booktitle = {The Thirty-Fourth {AAAI} Conference on Artificial Intelligence, {AAAI}
               2020, The Thirty-Second Innovative Applications of Artificial Intelligence
               Conference, {IAAI} 2020, The Tenth {AAAI} Symposium on Educational
               Advances in Artificial Intelligence, {EAAI} 2020, New York, NY, USA,
               February 7-12, 2020},
  pages     = {5436--5443},
  year      = {2020}
}

If you have any questions, please email to qiuzy@whu.edu.cn