Next Basket Recommendation with Time-Aware Item Weighting

This repository modified based on TAIW (Romanov, A., Lashinin, O., Ananyeva, M. and Kolesnikov, S., 2023, September. Time-Aware Item Weighting for the Next Basket Recommendations. In Proceedings of the 17th ACM Conference on Recommender Systems (pp. 985-992).) [pdf].)

Project Structure

.
├── EMOS // folder with experiments on EMOS dataset
│   ├── best_checkpoint-emos-*.pth // best pre-trained model from each period scenario
│   ├── best_checkpoint.pth // overall best pre-trained model
│   ├── Prep-emos.ipynb // notebook for preprocessing
│   ├── TAIWI-emos.ipynb // notebook for doing training and inference
│   └── TAIWI-emos-emb.ipynb // notebook for doing training and inference (with embedding scenario)
├── EPM // folder with experiments on EPM dataset
│   ├── best_checkpoint-epm-*.pth // best pre-trained model for each period scenario
│   ├── best_checkpoint.pth // overall best pre-trained model
│   ├── Prep-epm.ipynb // notebook for preprocessing
│   └── TAIWI-epm.ipynb // notebook for doing training and inference
├── nbr // NBR library
│   ├── common // subpackage with recommendation metrics definition and common constants
│   ├── dataset // subpackage with NBR dataset definition
│   ├── model // subpackage with NBR models definition
│   ├── preparation // subpackage with dataset preprocessing classes
│   ├── trainer // subpackage with NBR trainer definition
│   └── __init__.py
├── notebooks // folder with experiments on three real-world datasets
│   ├── data // folder with unprocessed datasets
│   ├   ├── dunnhumby.txt // Dunnhumby dataset
│   ├   ├── ta_feng.txt // TaFeng dataset
│   ├   └── taobao.txt // TaoBao dataset (purchase interactions only)
│   ├── testing // folder with testing DNNTSP model
│   ├   ├── DNNTSP // DNNTSP library (authors version with minor modifications)
│   └── └── testing_*.ipynb // notebook for testing various model
├── Tafeng // folder with experiments on Tafeng dataset
│   ├── best_checkpoint.pth // overall best pre-trained model
│   └── TAIWI-Tafeng.ipynb // notebook for doing training and inference
├── .gitignore // ignored files by git
├── README.md
├── requirements.txt // requirements for installing python dependencies
└── script.py // script for doing the training process automatically (equivalent to the notebook)

Preparation

Installation

# install requirements
pip install -r requirements.txt

Dataset (only for Local)

Download dataset that you want to use (EMOS, EPM, or Tafeng). Extract it in its respective directory so the structure will be look like this: (example for EMOS)

.
├── EMOS 
│   ├── data
│   ├   ├── data_EMOS-prep // generated directory 
│   ├   ├── EMOS-prep.txt // preprocessed data
│   ├   └── EMOS.csv // raw data
│   ├── Prep-emos.ipynb 
└── └── TAIWI-emos.ipynb 

Installation

# install requirements
pip install -r requirements.txt

Usage

Cloud (Kaggle)

In regard to the execution of our main and latest code, TAIW-emos-emb.ipynb, it is advisable to leverage cloud-based computing resources due to the program's substantial memory requirements. Furthermore, the embedding data is only accessible through Kaggle Datasets. By utilizing Kaggle Notebooks, the data can be directly incorporated into the analysis, streamlining the workflow.

Local

Preprocessing code can be found on Prep-*.ipynb (for EMOS and EPM). But it is optional because the preprocessed data already available in the downloaded dataset before.

Training code can be found on TAIWI-*.ipynb and will generate the best model as best-checkpoint.pth.

Training code also can be run by executing the script:

python script.py

Evaluation

So far, this is the best metrics we can achieve by performing hyperparameter tuning on each dataset, and its comparation with the original data:

Dataset	NDCG	Recall	Precision
EMOS	0.3214	0.3738	0.1629
EPM	0.2551	0.3414	0.1121
Tafeng*	0.1267	0.1642	0.0671
TaoBao*	0.0815	0.1190	0.0123
Dunnhumny*	0.1713	0.1791	0.1214

*Obtained from the original paper