README.md

WWW'21 HGCF: Hyperbolic Graph Convolution Networks for Collaborative Filtering

Authors: Jianing Sun*, Zhaoyue Cheng*, Saba Zuberi, Felipe Perez, Maksims Volkovs
[paper]

Environment:

The code was developed and tested on the following python environment:

python 3.7.7
pytorch 1.5.1
scikit-learn 0.23.2
numpy 1.19.1
scipy 1.5.4
tqdm 4.48.2

Instructions:

1. Preprocess the Amazon datasets (CDs and Vinyl, Books), Yelp2020 dataset, run this command:

python utils.preprocessing.py --dataset [Amazon_CD|Amazon_Book|yelp] --read_path your_raw_data_file 

Update: Original Yelp data on the website has been updated by Yelp in 2021, and the 2020 version we used has been overwritten. Using our preprocessing script on the updated Yelp data will give different statistics and thus different results. We've uploaded our preprocessed .pkl file of Yelp2020 here.

2. Train and evaluation HGCF:

• config.py for training on Amazon-CD dataset:

config_args = {
    'training_config': {
        'log': (None, 'None for no logging'),
        'lr': (0.001, 'learning rate'),
        'batch-size': (10000, 'batch size'),
        'epochs': (500, 'maximum number of epochs to train for'),
        'weight-decay': (0.005, 'l2 regularization strength'),
        'momentum': (0.95, 'momentum in optimizer'),
        'seed': (1234, 'seed for data split and training'),
        'log-freq': (1, 'how often to compute print train/val metrics (in epochs)'),
        'eval-freq': (20, 'how often to compute val metrics (in epochs)'),
    },
    'model_config': {
        'embedding_dim': (50, 'user item embedding dimension'),
        'scale': (0.1, 'scale for init'),
        'dim': (50, 'embedding dimension'),
        'network': ('resSumGCN', 'choice of StackGCNs, plainGCN, denseGCN, resSumGCN, resAddGCN'),
        'c': (1, 'hyperbolic radius, set to None for trainable curvature'),
        'num-layers': (4,  'number of hidden layers in encoder'),
        'margin': (0.1, 'margin value in the metric learning loss'),
    },
    'data_config': {
        'dataset': ('Amazon-CD', 'which dataset to use'),
        'num_neg': (1, 'number of negative samples'),
        'test_ratio': (0.2, 'proportion of test edges for link prediction'),
        'norm_adj': ('True', 'whether to row-normalize the adjacency matrix'),
    }
}

• config.py for training on Amazon-Book dataset:

config_args = {
    'training_config': {
        'log': (None, 'None for no logging'),
        'lr': (0.001, 'learning rate'),
        'batch-size': (10000, 'batch size'),
        'epochs': (500, 'maximum number of epochs to train for'),
        'weight-decay': (0.0005, 'l2 regularization strength'),
        'momentum': (0.95, 'momentum in optimizer'),
        'seed': (1234, 'seed for data split and training'),
        'log-freq': (1, 'how often to compute print train/val metrics (in epochs)'),
        'eval-freq': (20, 'how often to compute val metrics (in epochs)'),
    },
    'model_config': {
        'embedding_dim': (50, 'user item embedding dimension'),
        'scale': (0.1, 'scale for init'),
        'dim': (50, 'embedding dimension'),
        'network': ('resSumGCN', 'choice of StackGCNs, plainGCN, denseGCN, resSumGCN, resAddGCN'),
        'c': (1, 'hyperbolic radius, set to None for trainable curvature'),
        'num-layers': (4,  'number of hidden layers in encoder'),
        'margin': (0.1, 'margin value in the metric learning loss'),
    },
    'data_config': {
        'dataset': ('Amazon-Book', 'which dataset to use'),
        'num_neg': (1, 'number of negative samples'),
        'test_ratio': (0.2, 'proportion of test edges for link prediction'),
        'norm_adj': ('True', 'whether to row-normalize the adjacency matrix'),
    }
}

• config.py for training on yelp dataset:

config_args = {
    'training_config': {
        'log': (None, 'None for no logging'),
        'lr': (0.001, 'learning rate'),
        'batch-size': (10000, 'batch size'),
        'epochs': (500, 'maximum number of epochs to train for'),
        'weight-decay': (0.001, 'l2 regularization strength'),
        'momentum': (0.95, 'momentum in optimizer'),
        'seed': (1234, 'seed for data split and training'),
        'log-freq': (1, 'how often to compute print train/val metrics (in epochs)'),
        'eval-freq': (20, 'how often to compute val metrics (in epochs)'),
    },
    'model_config': {
        'embedding_dim': (50, 'user item embedding dimension'),
        'scale': (0.1, 'scale for init'),
        'dim': (50, 'embedding dimension'),
        'network': ('resSumGCN', 'choice of StackGCNs, plainGCN, denseGCN, resSumGCN, resAddGCN'),
        'c': (1, 'hyperbolic radius, set to None for trainable curvature'),
        'num-layers': (4,  'number of hidden layers in encoder'),
        'margin': (0.2, 'margin value in the metric learning loss'),
    },
    'data_config': {
        'dataset': ('yelp', 'which dataset to use'),
        'num_neg': (1, 'number of negative samples'),
        'test_ratio': (0.2, 'proportion of test edges for link prediction'),
        'norm_adj': ('True', 'whether to row-normalize the adjacency matrix'),
    }
}

Citation

If you find this code useful in your research, please cite the following paper:

@inproceedings{sun2021hgcf,
  title={HGCF: Hyperbolic Graph Convolution Networks for Collaborative Filtering},
  author={Jianing Sun, Zhaoyue Cheng, Saba Zuberi, Felipe Perez, Maksims Volkovs},
  booktitle={Proceedings of the International World Wide Web Conference},
  year={2021}
}