MODT4R: Multi-Objective Decision Transformer for Recommendations

Paper: Beyond Accuracy: Decision Transformers for Reward-Driven Multi-Objective Recommendations

---

📖 Overview

MODT4R implements a novel multi-objective decision transformer framework for sequential recommendation systems. This work leverages decision transformers to optimize multiple objectives simultaneously, including accuracy, diversity, and novelty in recommendations.

Main Scripts

• Training: MODT4R.py - Train the multi-objective decision transformer
• Evaluation: test_MODT4R.py - Evaluate trained models on test data

---

🚀 Quick Start

🛠️ Environment Setup

Prerequisites

• Python 3.8.5
• CUDA 10+ (for GPU support)
• Anaconda/Miniconda

Installation

# Create conda environment from yml file
conda env create -f conda_env.yml
conda activate decision-transformer-rec

---

Training

python MODT4R.py --dataset retail_rocket

Evaluation

python test_MODT4R.py --dataset retail_rocket

---

📁 Repository Structure

MODT4R_public/
├── models/                         # Model architectures
│   ├── MoTrans.py                  # Multi-objective Decision Transformer (main model)
│   ├── SASRecModules.py            # SASRec attention modules (User State Transformer)
│   ├── trajectory_gpt2.py          # GPT-2 backbone for trajectory modeling
│   └── model.py                    # Base model class
│
├── src/                             # Core utilities and functions
│   ├── evaluate_mo.py              # Multi-objective evaluation functions
│   └── util.py                     # Data processing and utility functions
│
├── MODT4R.py                       # 🎯 Main training script
├── test_MODT4R.py                  # 🧪 Main evaluation script
├── conda_env.yml                   # Conda environment configuration
└── README.md                       # This file

---

📊 Data Format

The data should contain the following fields:

• state: User interaction history (sequence of items)
• rewards: Reward signals (accuracy, novelty, diversity)
• len_state: Length of the sequence
• action: Next item to predict

---

🚀 Training

Train MODT4R

Use MODT4R.py for training the Multi-Objective Decision Transformer:

# Train on RetailRocket dataset
python MODT4R.py \
  --dataset retail_rocket \
  --hidden_factor 64 \
  --num_epochs 100 \
  --batch_size 256 \
  --lr 0.0001 \
  --embed_dim 128 \
  --n_layer 1 \
  --n_head 1 \
  --dropout 0.1 \
  --data_path /path/to/data/ \
  --models_path /path/to/save/models

---

🧪 Evaluation

Run Evaluation

Use test_MODT4R.py to evaluate trained MODT4R models:

# Evaluate model on RetailRocket dataset
python test_MODT4R.py \
  --dataset retail_rocket \
  --hidden_factor 64 \
  --embed_dim 128 \
  --n_layer 1 \
  --n_head 1 \
  --start_step 1 \
  --end_step 2

# Evaluate on RC15 dataset
python test_MODT4R.py \
  --dataset rc15 \
  --hidden_factor 64 \
  --embed_dim 128

Evaluation Metrics

The evaluation includes:

• Accuracy Metrics: Hit Rate (HR@K), NDCG@K
• Diversity Metrics: coverage
• Novelty Metrics: long-tail coverage

---

📝 Citation

@article{wang2025beyond,
  title={Beyond Accuracy: Decision Transformers for Reward-Driven Multi-Objective Recommendations},
  author={Wang, Jie and Karatzoglou, Alexandros and Arapakis, Ioannis and Jose, Joemon M and Ge, Xuri},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2025},
  publisher={IEEE}
}

---

🤝 Acknowledgments

This work builds upon several foundational works:

• Decision Transformer - Offline RL with transformers

---

---

📄 License

This project is released under the MIT License. See LICENSE file for details.

---