allenjack / hgn

Hierarchical Gating Networks for Sequential Recommendation

The HGN model for Sequential Recommendation

The implementation of the paper:

Chen Ma, Peng Kang, and Xue Liu, "Hierarchical Gating Networks for Sequential Recommendation", in the 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2019)

Arxiv: https://arxiv.org/abs/1906.09217

Please cite our paper if you use our code. Thanks!

Author: Chen Ma ([email protected])

Feel free to send me an email if you have any questions.

Bibtex

@inproceedings{DBLP:conf/kdd/MaKL19,
  author    = {Chen Ma and
               Peng Kang and
               Xue Liu},
  title     = {Hierarchical Gating Networks for Sequential Recommendation},
  booktitle = {{KDD}},
  pages     = {825--833},
  publisher = {{ACM}},
  year      = {2019}
}

Environments

• python 3.6
• PyTorch (version: 1.0.0)
• numpy (version: 1.15.0)
• scipy (version: 1.1.0)
• sklearn (version: 0.19.1)

Dataset

In our experiments, the movielens-20M dataset is from https://grouplens.org/datasets/movielens/20m/, the Amazon-CDs and Amazon-Books datasets are from http://jmcauley.ucsd.edu/data/amazon/, the GoodReads-Children and GoodReads-Comics datasets are from https://sites.google.com/eng.ucsd.edu/ucsdbookgraph/home. (If you need the data after preprocessing, please send me an email check this Google Drive link).

The XXX_tem_sequences.pkl file is a list of lists that stores the inner item id of each user in a chronological order, e.g., user_records[0]=[item_id0, item_id1, item_id2,...].

The XXX_user_mapping.pkl file is a list that maps the user inner id to its original id, e.g., user_mapping[0]=A2SUAM1J3GNN3B.

The XXX_item_mapping.pkl file is similar to XXX_user_mapping.pkl.

Example to run the code

Data preprocessing:

The code for data preprocessing is put in the /preprocessing folder. Amazon_CDs.ipynb provides an example on how to transform the raw data into the .pickle files that used in our program.

Train and evaluate the model (you are strongly recommended to run the program on a machine with GPU):

python run.py

Another implementation

Kindly thank Liwei Wu for another implementation with a different evaluation strategy: https://github.com/wuliwei9278/HGN_baseline.

Acknowledgment

The sequence segmentation (interactions.py) is heavily built on Spotlight. Thanks for the amazing work.

hgn's Issues

How to understand shift here?

In def data_index_shift(self, lists, increase_by=2) the explanation is:

Increase the item index to contain the pad_index

In def generate_dataset(self, index_shift=1), return self.num_items + index_shift

How to understant this? Any response is appreciated.

What does the meaning of this row ?

run.py

what's the meaning of this row ? Thanks for your response.

The experimental results are not match the paper results

Hello, I run your code provided, the result of Recall@10 is worse than the paper's, but the NDCG@10 is better than the paper's. And the NDCG@10 is big margin than paper's. Will you update your paper's results in the future?

How to predict New values?

The current model predicts values from the test set.
It there a way to pass our own values through HGN for prediction (from indexes) than creating another test set like that?

Example: How to generate the prediction for a user with their Books_item_Sequences = [427, 1347, 7598, 4876, 32334 ]?

Any help or guidance is appreciated.

A mistake on negative sampling

HGN/run.py

Line 91 in 8bf6008

return np.asarray(neg_samples).reshape(num_sets, train_matrix.shape[0], num_neg)

I think you may change it to this one. Otherwise, positive items can be sampled.

ret = np.asarray(neg_samples).reshape(train_matrix.shape[0], num_sets, num_neg)  # [n_user, ]
return np.transpose(ret, axes=[1, 0, 2])```

Would the parameters work if the output dimension of training and prediction are different ?

In the code(gating_network.py) above the output dimensions are different for training(batch_size,(target_length+neg_length)) and test(batch_size, num_item).
So my question is whether the parameters still work for test ? Is this a common operation for sequential recommendation ?
Thank you for your response !