ruizhang2016 / graphbasedrnn

Graph-Based Recursive Neural Network for Vertex Classification

GraphBasedRNN

Graph-Based Recursive Neural Network for Vertex Classification.

Usage

• Clone the repository

git clone https://github.com/xuqiongkai/GraphBasedRNN.git --recursive

• Prepare the dataset (optional)

Datasets used for the evaluation in our work has been processed to the supported format. You may also prepare your own dataset according to the Dataset format section.

• Run the desired models by main.lua script. To evaluate on a dataset, simply specify the folder path with its prefix

th main/main.lua -d  NoiseGraphDataset/citeseer <arguments>

Please refer to th main/main.lua -h for complete list of parameters.

Dataset format

Each dataset need to be assigned with a prefix. Then the folder structure and file names are determined by following rules:

├── prefix
│ ├── prefix.content
│ ├── prefix.feature
│ ├── prefix.cites
│ ├── prefix.label
│ ├── prefix.meta
│ |
│ ├── prefix.cites.add10
│ ├── prefix.cites.add5
│ ├── prefix.cites.reduce10
│ ├── prefix.cites.reduce5
│ |
│ └── README

Components

• prefix.label: parsed labels from content file, each row contains the tab separated entry ID and class label
• prefix.feature: serialized feature matrix in torch.DoubleTensor, 2-D matrix of dimension num_examples * number_of_fetures.
• prefix.cites: links in tab-separated format with each row denoting a single edge; the two elements separated by the tab are the two vertices connected by this link
• prefix.meta: dictionary of all possible class labels in this dataset

Noisy variants

• prefix.cites.add5: links data with additional 5% random links
• prefix.cites.add10: links data with additional 10% random links
• prefix.cites.reduce5: links data by randomly removing 5% links from the original
• prefix.cites.reduce10: links data by randomly removing 10% links from the original

Supported models

• Logistic Regression (LR)
• Iterative Classification Approach (ICA)
• Label Propagation (LP)
• Text-Associated DeepWalk (TADW)
• GRNN
• GRANN

Please refer to our paper cikm2017 for more details of the model.

Acknowledgement

It would be appreciated to cite the related publications (ordered by publication date) if you decide to use these datasets in your work.

Citation

TODO

Supplementary references

@article{mccallum2000automating,
    title={Automating the construction of internet portals with machine learning},
    author={McCallum, Andrew Kachites and Nigam, Kamal and Rennie, Jason and Seymore, Kristie},
    journal={Information Retrieval},
    volume={3},
    number={2},
    pages={127--163},
    year={2000},
    publisher={Springer}
}

@inproceedings{giles1998citeseer,
    title={CiteSeer: An automatic citation indexing system},
    author={Giles, C Lee and Bollacker, Kurt D and Lawrence, Steve},
    booktitle={Proceedings of the third ACM conference on Digital libraries},
    pages={89--98},
    year={1998},
    organization={ACM}
}

@inproceedings{Craven:1998:LES:295240.295725,
    author = {Craven, Mark and DiPasquo, Dan and Freitag, Dayne and McCallum, Andrew and Mitchell, Tom and Nigam, Kamal and Slattery, Se'{a}n},
    title = {Learning to Extract Symbolic Knowledge from the World Wide Web},
    booktitle = {Proceedings of the Fifteenth National/Tenth Conference on Artificial Intelligence/Innovative Applications of Artificial Intelligence},
    series = {AAAI '98/IAAI '98},
    year = {1998},
    location = {Madison, Wisconsin, USA},
    pages = {509--516},
    numpages = {8},
    publisher = {American Association for Artificial Intelligence},
    address = {Menlo Park, CA, USA},
}

License

The GNU General Public License v3.0