Skip to content

Latest commit

 

History

History
69 lines (54 loc) · 2.37 KB

README.md

File metadata and controls

69 lines (54 loc) · 2.37 KB
Raw

TAGNN

Paper & Dataset

This is the code for the SIGIR 2020 Paper: TAGNN: Target Attentive Graph Neural Networks for Session-based Recommendation.

I implemented with Dressip which is the dataset for RecSys Challenge 2022. Download the dressipi dataset and put it in datasets/.
The original used datasets in TAGNN are YOOCHOOSE (RecSys Challenge 2015) and DIGINETICA.

Notes:
Actually, the dataset I used is train_sessions_purchases.csv which is not in the official dressipi dataset.
This csv is the combination of train_sessions.csv and train_purchases.csv. I put each session in train_purchases.csv at the end of train_session.csv according to their session_id to generate train_sessions_purchases.csv (Thanks to my group member).

For example:
This

session_id,item_id,date
3,9655,2020-12-18 21:25:00.373
3,9655,2020-12-18 21:19:48.093

becomes

session_id,item_id,date
3,9655,2020-12-18 21:25:00.373
3,9655,2020-12-18 21:19:48.093
3,15085,2020-12-18 21:26:47.986 <- this is the purchase of this session

There is a small dataset sample included in the folder datasets/, which can be used to test the correctness of the code.

Requirments

  • Python 3
  • PyTorch 1.4.0

Usage

Dataset

You need to run the datasets/preprocess.py first to generate the correct data format.

python3 preprocess.py --dataset dressipi --date 2021-04 --cut

optional arguments:
--dataset: dataset name
--date: start date of training dataset
--cut: cut the test dataset

Tips:
The generated datasets are binary file, you can't read them.
Use this cool command, for example:
python3 -mpickle dressipi*/train.txt > dressipi*/train_txt.txt

Train

python3 main.py --dataset [dressipi*]

Results

My testing results (MRR@100):
I also tested on SR-GNN with the same dataset, and the performances of TAGNN are indeed a little bit better.

start month 2020-01 2021-01 2021-03 2021-04
SR-GNN 11.111 11.528 11.294
TAGNN 15.311 11.6 11.778 11.847

Others

Actually, the whole TAGNN code is based on SR-GNN (the research group basically consists of same people).
If you have any further questions, check it on the SR-GNN repo.