kunwuz / dgtn

[ICDMW 2020] Code and dataset for "DGTN: Dual-channel Graph Transition Network for Session-based Recommendation"

DGTN: Dual-channel Graph Transition Network for Session-based Recommendation

This repository contains PyTorch Implementation of ICDMW 2020 (NeuRec @ ICDM) paper: DGTN: Dual-channel Graph Transition Network for Session-based Recommendation. Please check our paper for more details about our work if you are interested.

Usage

Following the steps below to run our codes:

1. Preprocess

The preprocess code is in preprocess/

2. Neighbors retrieval

Please run neigh_retrieval/neighborhood_retrieval.py

3. Run the model

Please run main.py

Requirements

• Python 3
• PyTorch 1.1.0

Citation

If you find this repo is useful for you, please kindly cite our paper.

@inproceedings{zheng2020dgtn,
    title={DGTN: Dual-channel Graph Transition Network for Session-based Recommendation},
    author={Zheng, Yujia and Liu, Siyi and Li, Zekun and Wu, Shu},
    booktitle={ICDMW},
    year={2020},
}

dgtn's Issues

Dataset

Could you give me the preprocessed result of Diginetica? Thank you very much.

Error ocurred when running neighborhood_retrieval.py

Hi, I follow the usage in the readme and an error occurred when I run neighborhood_retrieval.py using the “Diginetica” dataset.
The console shows that ：
File "..\DGTN-master\neigh_retrieval\knn.py", line 58, in find_sess
sess_index += item_sess_map[item]
KeyError: 31938

I think the error is resulted from the function get_item_sess_map in knn.py line 19:
items = np.unique(sess[:-1])
because the item_id:31938 may be the last click item in the session, and it is not the recorded in the item_sess_map
I wonder if it is reasonable to change this line into "items = np.unique(sess) " to solve the problem?

looking forward to u reply, thank u!

experiment

hi.
I cannot achieve the results in your paper using Diginetica data set. Are there any other parameters that need to be set?

experiment

Hi,
Under the convolution fusion function, when running the program with different neighborhood files, the experimental results did not change, but when running under the max function, the results changed but did not reach the results in your paper.

Is the neighborhood data generated by me wrong? Looking forward to your reply . Thank you very much!

What is the experimental environment？

What is data structure of dataset/train.txt ?

Hi all,
I don't know the schema of dataset files.
Can you give me a description?

Hi, is this an error?

DGTN/model/multi_sess.py

Line 116 in 4f3d668

hidden - self.sgcn(hidden, edge_index)