# OpenNE: An open source toolkit for Network Embedding
This repository provides a standard NE/NRL(Network Representation Learning）training and testing framework. In this framework, we unify the input and output interfaces of different NE models and provide scalable options for each model. Moreover, we implement typical NE models under this framework based on tensorflow, which enables these models to be trained with GPUs.
We develop this toolkit according to the settings of DeepWalk. The implemented or modified models include [DeepWalk](https://github.com/phanein/deepwalk), [LINE](https://github.com/tangjianpku/LINE), [node2vec](https://github.com/aditya-grover/node2vec), [GraRep](https://github.com/ShelsonCao/GraRep), [TADW](https://github.com/thunlp/TADW) and [GCN](https://github.com/tkipf/gcn). We will implement more representative NE models continuously according to our released [NRL paper list](https://github.com/thunlp/nrlpapers). Specifically, we welcome other researchers to contribute NE models into this toolkit based on our framework. We will announce the contribution in this project.
## Requirements
- numpy==1.13.1
- networkx==2.0
- scipy==0.19.1
- tensorflow==1.3.0
- gensim==3.0.1
- scikit-learn==0.19.0
## Usage
#### General Options
You can check out the other options available to use with *OpenNE* using:
python src/main.py --help
- --input, the input file of a network;
- --graph-format, the format of input graph, adjlist or edgelist;
- --output, the output file of representation (GCN doesn't need it);
- --representation-size, the number of latent dimensions to learn for each node; the default is 128
- --method, the NE model to learn, including deepwalk, line, node2vec, grarep, tadw and gcn;
- --directed, treat the graph as directed; this is an action;
- --weighted, treat the graph as weighted; this is an action;
- --label-file, the file of node label; ignore this option if not testing;
- --clf-ratio, the ratio of training data for node classification; the default is 0.5;
- --epochs, the training epochs of LINE and GCN; the default is 5;
#### Example
To run "node2vec" on BlogCatalog network and evaluate the learned representations on multi-label node classification task, run the following command in the home directory of this project:
To run "gcn" on Cora network and evaluate the learned representations on multi-label node classification task, run the following command in the home directory of this project:
- --number-walks, the number of random walks to start at each node; the default is 10;
- --walk-length, the length of random walk started at each node; the default is 80;
- --workers, the number of parallel processes; the default is 8;
- --window-size, the window size of skip-gram model; the default is 10;
- --q, only for node2vec; the default is 1.0;
- --p, only for node2vec; the default is 1.0;
LINE:
- --negative-ratio, the default is 5;
- --order, 1 for the 1st-order, 2 for the 2nd-order and 3 for 1st + 2nd; the default is 3;
- --no-auto-save, no early save when training LINE; this is an action; when training LINE, we will calculate F1 scores every epoch. If current F1 is the best F1, the embeddings will be saved.
GraRep:
- --kstep, use k-step transition probability matrix（make sure representation-size%k-step == 0).
TADW:
- --lamb, lamb is a hyperparameter in TADW that controls the weight of regularization terms.
GCN:
- --feature-file, The file of node features;
- --epochs, the training epochs of GCN; the default is 5;
- --dropout, dropout rate;
- --weight-decay, weight for l2-loss of embedding matrix;
- --hidden, number of units in the first hidden layer.
#### Input
The supported input format is an edgelist or an adjlist:
edgelist: node1 node2 <weight_float,optional>
adjlist: node n1 n2 n3 ... nk
The graph is assumed to be undirected and unweighted by default. These options can be changed by setting the appropriate flags.
If the model needs additional features, the supported feature input format is as follow (**feature_i** should be a float number):
node feature_1 feature_2 ... feature_n
#### Output
The output file has *n+1* lines for a graph with *n* nodes.
The first line has the following format:
num_of_nodes dim_of_representation
The next *n* lines are as follows:
node_id dim1 dim2 ... dimd
where dim1, ... , dimd is the *d*-dimensional representation learned by *OpenNE*.
#### Evaluation
If you want to evaluate the learned node representations, you can input the node labels. It will use a portion (default: 50%) of nodes to train a classifier and calculate F1-score on the rest dataset.
The supported input label format is
node label1 label2 label3...
## Comparisons with other implementations
Running environment: <br/>
BlogCatalog: CPU: Intel(R) Xeon(R) CPU E5-2620 v3 @ 2.40GHz. <br/>
Wiki, Cora: CPU: Intel(R) Core(TM) i5-7267U CPU @ 3.10GHz. <br/>
We show the node classification results of various methods in different datasets. We set representation dimension to 128, **kstep=4** in GraRep.
Note that, both GCN(a semi-supervised NE model) and TADW need additional text features as inputs. Thus, we evaluate these two models on Cora in which each node has text information. We use 10% labeled data to train GCN.
[Wiki](https://github.com/thunlp/MMDW/tree/master/data) (Wiki dataset is provided by [LBC project](http://www.cs.umd.edu/~sen/lbc-proj/LBC.html). But the original link failed.): 2405 nodes, 17981 edges, 19 labels, directed:
