MARS (CIKM'24)

This is an official repository for our paper "MARS: Matching Attribute-aware Representations for Text-based Sequential Recommendation" in CIKM'24.

Overview

MARS is a text-based sequential recommendation framework that effectively captures attribute-wise user/item interactions.

• Attribute-aware text encoding captures the fine-grained user preferences based on textual attributes of items.
• Attribute-wise interaction matching identifies the attribute-level preference of users.

Please refer to the paper and poster for more details

Paper: HERE

Poster: HERE

Getting Started

Environment

Please refer to the requirements.txt file for the required packages.

pytorch-lightning==2.3.3
transformers~=4.28.0
wandb
wonderwords

Dataset

Dataset can downloaded from HERE. Please download the 5-core dataset and metadata, and unzip it to the dataset folder. Run process.py as follows:

python process.py --file_path path/to/dataset.json.gz --meta_file_path path/to/meta_dataset.json.gz --output_path dataset_name

Training

Run the training script as follows:

python main.py --data_path dataset/Scientific_ours --bf16 --num_train_epochs 128 --warmup_steps 800

Acknowledgement

This work is based on and inspired by the methods introduced in Recformer.

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{kim2024mars,
  title={MARS: Matching Attribute-aware Representations for Text-based Sequential Recommendation},
  author={Kim, Hyunsoo and Kim, Junyoung and Choi, Minjin and Lee, Sunkyung and Lee, Jongwuk},
  booktitle={Proceedings of the 33rd ACM International Conference on Information and Knowledge Management},
  pages={3822--3826},
  year={2024}
}