This project implements the E-net model, which focuses on reconstructing a reliable network from a flawed (noisy) network.
The script has been tested running under Python 2.7.12, with the following packages installed (along with their dependencies):
torch==1.1.0networkx==2.0sklearn==0.19.1numpy==1.11.0scipy==1.1.0gensim==3.6.0tqdm==4.19.4
Some Python module dependencies are listed in requirements.txt, which can be easily installed with pip:
pip install -r requirements.txt
In addition, CUDA 8.0 has been used in our project. Although not all dependencies are mentioned in the installation instruction links above, you can find most of the libraries in the package repository of a regular Linux distribution.
An example data format is given in Enet/data where dataset is in mat format and score is some heuristic scores calculated in advance. Besides, Enet/embedding(OPTIONAL) provides additional node embedding results by Node2vec (https://github.com/aditya-grover/node2vec).
When using your own dataset, you must provide:
- an N by N adjacency matrix (N is the number of nodes).
- an N by D feature matrix (D is the dimension of node features).
- an N by C one-hot label matrix (C is the number of classes).
- an N by K heuristic score matrix (K is the number of heuristic scores).
The help information of the main script Enet/Main.py is listed as follows:
python Main.py -h
usage: Main.py [-h][--data-name] [--save-name] [--max-train-num] [--no-cuda] [--missing-ratio]
[--split-ratio] [--neg-pos-ratio] [--use-attribute] [--use-embedding] [--embedding-size]
[--lazy-subgraph] [--max-nodes-per-hop] [--num-walks] [--multi-subgraph] [--reg-smooth]
[--smooth-coef] [--trainable-noise] [--early-stop] [--early-stop-patience] [--learning-rate]
optional arguments:
-h, --help show this help message and exit
--data-name str, select the dataset.
--save-name str, the name of saved model.
--max-train-num int, the maximum number of training links.
--no-cuda bool, whether to disables CUDA training.
--seed int, set the random seed.
--test-ratio float, the ratio of test links.
--missing-ratio float, the ratio of missing links.
--split-ratio str, the split rate of train, val and test links
--neg-pos-ratio float, the ratio of negative/positive links
--use-attribute bool, whether to utilize node attribute.
--use-embedding bool, whether to utilize the information from node2vec node embeddings.
--embedding-size int, the embedding size of node2vec
--lazy-subgraph bool, whether to use lazy subgraph extraction.
--max-nodes-per-hop int, the upper bound the number of nodes per hop when performing Lazy Subgraph Extraction.
--num-walks int, thenumber of walks for each node when performing Lazy Subgraph Extraction.
--multi-subgraph int, the number of subgraphs to extract for each queried nodes
--reg-smooth bool, whether to use auxiliary denoising regularization.
--smooth-coef float, the coefficient of auxiliary denoising regularization.
--trainable-noise bool, whether to let the Noisy link detection layer trainable.
--early-stop bool, whether to use early stopping.
--early-stop-patience int, the patience for early stop.
--learning-rate float, the learning rate.
To reproduce the results that reported in the paper, you can run the following command:
python Main.py --data-name citeseer --use-embedding --num-walks 5 --learning-rate 1e-4
--noise-hidden-dim 500 --use-sig --use-soft --reg-smooth --smooth-coef 1e-4 --trainable-noise
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent