Enhancing Time Awareness in Generative Recommendation (EMNLP'25 Findings)

This is the official implementation for the paper: "Enhancing Time Awareness in Generative Recommendation".

🔍 Overview

GRUT (Generative Recommender Using Time awareness) is a novel generative recommendation system that effectively captures hidden user preferences via various temporal signals. It is built on top of GRAM (ACL'25), extending its semantic-aware multi-granular late fusion with time-aware prompting and trend-aware inference. The repository ships both pipelines, so GRAM and GRUT can be trained and evaluated side-by-side from a single codebase.

📖 Resources

• 📄 Paper: ACL Anthology | arXiv
• 🎨 Presentation: Poster | Slides
• 📝 Blog Post: Korean

🛠️ Environment Setup

Requirements

• Python: 3.9+
• PyTorch: 1.11.0
• Transformers: 4.26.0
• CUDA: 11.3

Installation

1. Clone the repository

git clone https://github.com/skleee/GRUT.git
cd GRUT

2. Create conda environment

conda create -n grut python=3.9
conda activate grut

3. Install dependencies

# Install general dependencies
pip install -r requirements.txt

# Install PyTorch with CUDA support
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 \
    --extra-index-url https://download.pytorch.org/whl/cu113

📊 Datasets

We evaluate GRUT on 4 datasets from two domains:

Amazon Review Datasets

• Product reviews from Amazon with rich metadata
  • Beauty: Cosmetics and personal care products
  • Toys: Children's toys and gaming products
  • Sports: Sports equipment and outdoor gear

Yelp Dataset

• Yelp: Local business reviews and ratings

Data Sources

• Amazon Review: Official Dataset
• Yelp: Official Dataset
• Preprocessed Data: Available in rec_datasets/ directory

For more details on the dataset structure, please refer to README.md in the rec_datasets/ directory.

🚀 Training

Execute the training scripts located in the command/ folder. Each script bakes in the paper-final hyperparameters for its dataset:

# Amazon Beauty dataset
bash train_grut_beauty.sh

# Amazon Toys dataset
bash train_grut_toys.sh

# Amazon Sports dataset
bash train_grut_sports.sh

# Yelp dataset
bash train_grut_yelp.sh

Trend-aware Inference

GRUT's trend-aware reranking is a two-step inference workflow.

1. Tune the trend mixing weight λ on the validation set

bash test_grut_tune_valid.sh -d Beauty -e 30 -g 0 -c <CHECKPOINT_DIR>

This grid-searches λ ∈ {0.1, 0.2, …, 1.0}. The first λ runs full beam search and caches the predictions to disk; subsequent λ values reuse the cache and only redo the (fast) score-adjustment step. Pick the λ that maximizes validation NDCG@5.

2. Final test evaluation with the chosen λ

bash test_grut_test.sh -d Beauty -e 30 -g 0 -l 0.3 -c <CHECKPOINT_DIR>

-l 0.3 is the chosen λ. Test-set predictions are cached for reuse.

<CHECKPOINT_DIR> points to the directory containing model_rec_phase_1_epoch_<EPOCH>.pt, e.g. log/Beauty/<run_id>/id_0_rec_30/.

🙏 Acknowledgments

This work builds upon several open-source projects:

• GRAM: Generative Recommendation via Semantic-aware Multi-granular Late Fusion
• IDGenRec: Generative recommendation framework
• OpenP5: Open-source P5 implementation
• FiD: Fusion-in-Decoder architecture

📜 Citation

If you find this work helpful, please consider citing our paper:

@inproceedings{lee2025grut,
  title     = {Enhancing Time Awareness in Generative Recommendation},
  author    = {Sunkyung Lee and Seongmin Park and Jonghyo Kim and Mincheol Yoon and Jongwuk Lee},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2025},
  pages     = {23917--23933},
  year      = {2025},
  url       = {https://aclanthology.org/2025.findings-emnlp.1300/}
}