MAERec

This is the PyTorch implementation by @HelloElwin for MAERec proposed in the paper Graph Masked Autoencoder for Sequential Recommendation published in SIGIR'23 by Yaowen Ye, Lianghao Xia, and Chao Huang.

1. Introduction

MAERec is a simple yet effective graph masked autoencoder that adaptively and dynamically distills global item transitional information for self-supervised augmentation through a novel adaptive transition path masking strategy. It naturally addresses the data scarcity and noise perturbation problems in sequential recommendation scenarios and avoids issues in most contrastive learning-based methods.

2. Environment

We suggest the following environment for running MAERec:

python==3.8.13
pytorch==1.12.1
numpy==1.18.1

3. How to run

Please first unzip the desired dataset in the dataset folder, and then run

• Amazon Books: python main.py --data books
• Amazon Toys: python main.py --data toys
• Retailrocket: python main.py --data retailrocket

More explanation of model hyper-parameters can be found here.

4. Running on customized datasets

The dataset should be structured into four files for MAERec:

• seq a list of user sequences for training
• tst a list of user sequences for testing
• neg the negative samples for each test sequence
• trn the i-i graph, which can be generated using the following code:

def construct_graphs(num_items=54756, distance=3, path='./books/'):
    with open(path + 'seq', 'rb') as fs:
        seq = pickle.load(fs)
    user = list()
    r, c, d = list(), list(), list()
    for i, seq in enumerate(seqs):
        print(f"Processing {i}/{len(seqs)}          ", end='\r')
        for dist in range(1, distance + 1):
            if dist >= len(seq): break;
            r += copy.deepcopy(seq[+dist:])
            c += copy.deepcopy(seq[:-dist])
            r += copy.deepcopy(seq[:-dist])
            c += copy.deepcopy(seq[+dist:])
    d = np.ones_like(r)
    iigraph = csr_matrix((d, (r, c)), shape=(num_items, num_items))
    print('Constructed i-i graph, density=%.6f' % (len(d) / (num_items ** 2)))
    with open(prefix + 'trn', 'wb') as fs:
        pickle.dump(iigraph, fs)

5. Citing information

If you find this work helpful to your research, please kindly consider citing our paper:

@inproceedings{ye2023graph,
  title={Graph Masked Autoencoder for Sequential Recommendation},
  author={Ye, Yaowen and Xia, Lianghao and Huang, Chao},
  booktitle={Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'23), July 23-27, 2023, Taipei, Taiwa},
  year={2023}
}