Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/zw-zhang/randne
This is a sample implementation of "Billion-scale Network Embedding with Iterative Random Projection" (ICDM 2018).
https://github.com/zw-zhang/randne
Last synced: about 1 month ago
JSON representation
This is a sample implementation of "Billion-scale Network Embedding with Iterative Random Projection" (ICDM 2018).
- Host: GitHub
- URL: https://github.com/zw-zhang/randne
- Owner: ZW-ZHANG
- Created: 2018-08-29T07:47:28.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2020-08-05T08:06:56.000Z (over 4 years ago)
- Last Synced: 2023-09-18T09:39:37.947Z (about 1 year ago)
- Language: Python
- Homepage:
- Size: 18.8 MB
- Stars: 17
- Watchers: 2
- Forks: 2
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# RandNE
This is the official implementation of "[Billion-scale Network Embedding with Iterative Random Projection](https://zw-zhang.github.io/files/2018_ICDM_RandNE.pdf)" (ICDM 2018).
We provide two implementations: MATLAB and Python. Note that the MATLAB version is used in producing original results in the paper.
### Requirements
```
MATLAB R2017a
or
Python >= 3.5.2
numpy >= 1.14.2
scipy >= 1.1.0
pandas >= 0.22.0
sklearn >= 0.0
```
### Usage
##### Example Usage
```
See Sample_Run.m or Sample_Run.py for a sample run on the BlogCatalog network
```
##### Functions
```
Sample_Run: a sample run on the BlogCatalog network, see annotations for detail
RandNE_Projection: calculate projection results of different orders by performing iterative random projection
Inputs: sparse adjacency matrix, order q, dimension d, whether use orthogonal projection, random seed
Outputs: a cell containing decomposed parts U_0 ... U_q
RandNE_Combine: combine different orders with given weights
Inputs: a list of decomposed parts generated by RandNE_Projection, a vector of weights for each order w_0 ... w_q
Outputs: final embedding vectors U
RandNE_Update: update embeddings for dynamic networks
Inputs: original adjacency matrix, the changes of adjacency matrix, original decomposed parts, whether use orthogonal projection, random seed
Outputs: updated decomposed parts
GS: perform Gram–Schmidt process to obtain orthogonal projections
```
##### A Note on Hyper-parameters
```
We find empirically that our method is somewhat sensitive to hyper-parameters (adjacency matrix/transition matrix, order q and weights for different orders w).
A recommended way for tuning hyper-parameters is to use grid search on cross-validation sets, which should be efficient since order and weights only affect the last step of our algorithm (RandNE_Combine).
The range for searching is suggested as follows: order is from 1 to 3 and weights can be searched by w_{i+1} = beta_i w_i where beta_i is from {0.01,0.1,1,10,100}.
A rule of thumb is that adjacency matrix should have decreasing weights (i.e. beta_i <= 1) and transition matrix should have increasing weights (i.e. beta_i >= 1).
The hyper-parameters used in our paper for BlogCatalog dataset is listed as a reference:
Reconstruction: adjacency matrix, q = 1, weights = [1, 0.1]
Link Prediction: adjaecncy matrix, q = 2, weights = [1, 1, 0.01]
Node Classification: transition matrix, q = 3, weights = [1, 100, 10000, 100000]
```
### Cite
If you find this code useful, please cite our paper:
```
@inproceedings{zhang2018billion,
title={Billion-scale Network Embedding with Iterative Random Projection},
author={Zhang, Ziwei and Cui, Peng and Li, Haoyang and Wang, Xiao and Zhu, Wenwu},
booktitle={2018 IEEE International Conference on Data Mining (ICDM)},
pages={787--796},
year={2018},
organization={IEEE}
}
```