The code and dataset for our paper: Personalizing Graph Neural Networks with Attention Mechanism for Session-based Recommendation (https://arxiv.org/abs/1910.08887), which has been submitted to TKDE. We have implemented our methods in Tensorflow.
Here are two datasets we used in our paper.
- Xing http://2016.recsyschallenge.com/
- Reddit https://www.kaggle.com/colemaclean/subreddit-interactions
The processed data can be downloaded: https://www.dropbox.com/sh/hwx2347ir1worag/AABJK6IBXHNBlbvrvKqw94YKa?dl=0
You need to run the file record.py first to preprocess the data to generate the tf.record formart data for training and test.
For example:
python record.py --dataset=all_data --data=xing --adj=adj_all --max_session=50. This will creat xing/ in datasets/.
This code can be given the following command-line arguments:
--dataset: choose to use fully data or samples, if all_data: use fully data, if None or sample: use sample data.
--data: the name of data set, we can choose xing or reddit.
--graph: graph neural network, default set is ggnn.
--max_session: the maximum length of historical sessions.
--max_length: the maximum length of current session.
--last: if True, the last one for testing, else, the next 20% for testing.
Then you can run the file train_last.py to train the model and test.
For example: python train_last.py --data=xing --mode=transformer --user_ --adj=all_adj --dataset=all_data --hiddenSize=100
This code can be given the following command-line arguments:
--dataset: choose to use fully data or samples, if all_data: use fully data, if None or sample: use sample data.
--data: the name of data set, we can choose xing or reddit.
--user_: whether to user user embedding.
--max_session: the maximum length of historical sessions.
--max_length: the maximum length of current session.
--adj: if adj_all, use normalized weights for adjacency matrices, else, use binary adjacency matrix.
--batchSize: batchsize.
--epoch: epoch.
--lr: learning rate.
--buffer_size: the maximum number of elements that will be added to the buffer. For details, see the use of tf.record, for Xing, we set 200000, Reddit is 100000.
- Python 3.6.5
- Tensorflow-gpu 1.10.0
@article{wu2019personalizing,
title={Personalizing Graph Neural Networks with Attention Mechanism for Session-based Recommendation},
author={Wu, Shu and Zhang, Mengqi and Jiang, Xin and Ke, Xu and Wang, Liang},
journal={arXiv preprint arXiv:1910.08887},
year={2019}
}