path-infomax's Introduction

Path-InfoMax

This repository holds the code used in our IJCAI-21 paper Unsupervised Path Representation Learning with Curriculum Negative Sampling.

Overview

Here we give a PyTorch implementation of PIM. The repository is organized as follows:

data/ includes training data sample, when use your own data you can follow this format;
models/ contains the implementation of the PIM pipeline (pim.py);
layers/ contains the implementation of the bilinear discriminator (discriminator.py);
utils/ contains the necessary processing tool (process.py).

To better understand the code, we recommend that you could read the code of DGI/Petar (https://arxiv.org/abs/1809.10341) in advance. Besides, you could further optimize the code based on your own needs.

Dataset

We use two datasets in our paper: Aalborg and Harbin and you can download the Harbin dataset to train the PIM, where the pre-processing is needed.

Requirements

Ubuntu OS (18.04)
PyTorch =1.7.1
Numpy >= 1.16.2
Pickle

Please refer to the source code to install the required packages in Python.

Usage

python train.py

Cite

Please cite our paper if you make advantage of PIM in your research:

@inproceedings{
IJCAI21,
title="{Unsupervised Path Representation Learning with Curriculum Negative Sampling}",
author={Sean Bin Yang, Jilin Hu, Chenjuan Guo, Jian Tang and Bin Yang},
booktitle={Proceedings of The  30th International Joint Conference on Artificial Intelligence},
year={2021},
}

Recommend Projects