A PyTorch implementation of LESSR (Lossless Edge-order preserving aggregation and Shortcut graph attention for Session-based Recommendation) from the paper:
Handling Information Loss of Graph Neural Networks for Session-based Recommendation, Tianwen Chen and Raymong Chi-Wing Wong, KDD '20
- PyTorch 1.6.0
- NumPy 1.19.1
- Pandas 1.1.3
- DGL 0.5.2
-
Install the requirements.
If you use Anaconda, you can create a conda environment with the required packages using the following command.conda env create -f packages.yml
Activate the created conda environment.
conda activate lessr -
Download and extract the datasets.
-
Preprocess the datasets using preprocess.py.
For example, to preprocess the Diginetica dataset, extract the file train-item-views.csv to the folderdatasets/and run the following command:python preprocess.py -d diginetica -f datasets/train-item-views.csv
The preprocessed dataset is stored in the folder
datasets/diginetica.
You can see the detailed usage ofpreprocess.pyby running the following command:python preprocess.py -h
-
Train the model using main.py.
If no arguments are passed tomain.py, it will train a model using a sample dataset with default hyperparameters.python main.py
The commands to train LESSR with suggested hyperparameters on different datasets are as follows:
python main.py --dataset-dir datasets/diginetica --embedding-dim 32 --num-layers 4 python main.py --dataset-dir datasets/gowalla --embedding-dim 64 --num-layers 4 python main.py --dataset-dir datasets/lastfm --embedding-dim 128 --num-layers 4
You can see the detailed usage of
main.pyby running the following command:python main.py -h
-
Use your own dataset.
- Create a subfolder in the
datasets/folder. - The subfolder should contain the following 3 files.
num_items.txt: This file contains a single integer which is the number of items in the dataset.train.txt: This file contains all the training sessions.test.txt: This file contains all the test sessions.
- Each line of
train.txtandtest.txtrepresents a session, which is a list of item IDs separated by commas. Note the item IDs must be in the range of[0, num_items). - See the folder datasets/sample for an example of a dataset.
- Create a subfolder in the
If you use our code in your research, please cite our paper:
@inproceedings{chen2020lessr,
title="Handling Information Loss of Graph Neural Networks for Session-based Recommendation",
author="Tianwen {Chen} and Raymond Chi-Wing {Wong}",
booktitle="Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '20)",
pages="1172-–1180",
year="2020"
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent