A Python implementation of the more flexible version of the method proposed in "Social trust Network Embedding", Pinghua Xu, Wenbin Hu, Jia Wu, Weiwei Liu, Bo Du and Jian Yang, ICDM 2019. In this implementation, one can use the co-occurance node pair as positive samples instead of pairwise edges.
This repository is organized as follows:
input/contains four example graphsWikiEditorWikiElecWikiRfaSlashdotEpinions;output/is the directory to store the learned node embeddings;src/contains the implementation of the proposed STNE model.
The implementation is tested under Python 3.7, with the folowing packages installed:
networkx==2.3numpy==1.16.5scikit-learn==0.21.3texttable==1.6.2tqdm==4.36.1
We investigated social trust network, which can be represented by a directed signed (un)weighted graph, in this work.
The code takes an input graph in .txt format. Every row indicates an edge between two nodes separated by a space or \t. The file does not contain a header. Nodes can be indexed starting with any non-negative number. Five example graphs are included in the input/ directory. Among these graphs, WikiElec, WikiRfa, Slashdot and Epinions are donwloaded from SNAP, but node ID is resorted, and WikiEditor is generated according to the description in our paper. The structure of the input file is the following:
| Source node | Target node | Weight |
|---|---|---|
| 0 | 1 | 4 |
| 1 | 3 | -5 |
| 1 | 2 | 2 |
| 2 | 4 | -1 |
--edge-path STR Input file path Default=="./input/WikiElec.txt"
--outward-embedding-path STR Outward embedding path Default=="./output/WikiElec_outward"
--inward-embedding-path STR Inward embedding path Default=="./output/WikiElec_inward"
--dim INT Dimension of latent factor vector Default==32
--n INT Number of noise samples Default==5
--num_walks INT Walks per node Default==20
--walk_len INT Length per walk Default==10
--workers INT Number of threads used for random walking Default==4
--m FLOAT Damping factor Default==1
--norm FLOAT Normalization factor Default==0.01
--learning-rate FLOAT Leaning rate Default==0.02
--test-size FLOAT Test ratio Default==0.2
--split-seed INT Random seed for splitting dataset Default==16
Train an STNE model on the deafult WikiElec dataset, output the performance on sign prediction task, and save the embeddings:
python src/main.py
Train an SLF model with custom split seed and test ratio:
python src/main.py --split_seed 20 --test-size 0.3
Train an SLF model on the WikiRfa dataset:
python src/main.py --edge-path ./input/WikiRfa.txt --outward-embedding-path ./output/WikiRfa_outward --inward-embedding-path ./output/WikiRfa_inward
We perform sign prediction to evaluate the node embeddings. And we use AUC and macro-F1 as evaluation metric. Although Micro-F1 is used in our paper, we admit that it is not a good choice for evaluation on a dataset with unbalanced labels.
Run python src/main.py, and the output is printed like the following:
Optimizing: 100%|█████████████████████████████████████████████| 75496/75496 [13:20<00:00, 94.32it/s]
Sign prediction: AUC 0.874, F1 0.747
Like the other methods in Skip-Gram family, we only perform one-epoch training.
In our paper, we used the following methods for comparison:
N2V"node2vec: Scalable Feature Learning for Networks" [source]LINE"LINE: Large-scale information network embedding" [source]LSNE"Solving link-oriented tasks in signed network via an embedding approach"MF"Low rank modeling of signed networks"SIDE"Side: representation learning in signed directed networks" [source]SIGNet"Signet: Scalable embeddings for signed networks" [source]
If you find this repository useful in your research, please cite our paper:
@INPROCEEDINGS{8970926,
author={Pinghua Xu and Wenbin Hu and Jia Wu and Weiwei Liu and Bo Du and Jian Yang},
booktitle={2019 IEEE International Conference on Data Mining (ICDM)},
title={Social Trust Network Embedding},
year={2019},
pages={678-687}
}
Moreover, if you are interested in the topic of social trust network, you may want to know our another work "Opinion Maximization in Social Trust Networks" (IJCAI 2020).
Correction on a typo. Page 7, Sec. IV.A. "In the relationship prediction tasks, we randomly selected \alpha % of the relationships as the test set, and the remaining (100 - \alpha) % were used for training" -> "In the relationship prediction tasks, we randomly selected \alpha % of the relationships as the training set, and the remaining (100 - \alpha) % were used for test"