🚀 Adaptive Budget Optimization for Multichannel Advertising Using Combinatorial Bandits (AAMAS 2025 - Extended Abstract)

Welcome to the Adaptive Budget Optimization repository! This project focuses on simulating long-running ad campaigns and implementing the TUCBMAE algorithm to optimize budget allocation efficiently. 📊📈 It is the official implementation of the paper by Briti Gangopadhyay, Zhao Wang, Alberto Silvia Chiappa and Takamatsu Shingo.

---

📌 System Requirements

The code has been successfully tested on:

• 🖥 Ubuntu 20.04.2 LTS

---

📂 Datasets

This repository includes two real-world advertising datasets, each with three sub-campaigns run on the Google Ads platform:

1. 📡 Attendance System
2. 🌐 Internet Service Provider

🔹 Special thanks to Sony Biz Networks Corporation for providing this dataset from their services NURO Biz and AKASHI.

🔐 Privacy Note: Click and cost values have been randomly projected to maintain predictive power while ensuring that original values cannot be reconstructed. The Criteo dataset results are fully reproducible.

---

⚙️ Installation & Dependencies

Follow these steps to set up the required environment:

conda create --name tucbmae python=3.8
conda activate tucbmae
python -m pip install -r requirements.txt
pip install -e .

---

🔧 Hyperparameter Configuration

Configure hyperparameters in configpolicy.py. Key parameters include:

🔹 Simulation Parameters:

• num_month → Number of months for simulation (e.g., 16 for AI Prediction, 1 for Criteo).

🔹 Data Paths:

• data_google → Path to Google campaign data (e.g., data/attendance_campaign_data.csv).
• data_smn → Path to another platform’s data.
• data_criterio → Path to Criteo dataset (e.g., data/criterio_data_filtered.csv).

🔹 Experiment Settings:

• use_wandb → Enable logging results to Weights & Biases.
• baseline_name → Use GPUCBBaseline_Policy (GP-based) or Human (manual allocation).
• exploration_strategy → Choose ucb (UCB exploration) or ts (Thompson Sampling).
• adaptation_strategy → Select from discounted_reward, sliding_window, no_change_detection, mae_test.
• use_psudo_conversion → Use pseudo conversions as reward if True, else use clicks.

---

🎮 Simulation Environment

🔹 Key Features:

• 📅 Episodes → Each episode represents one month of ad campaign simulation.
• 💰 Actions → Budget allocation for campaigns.
• 🎯 Rewards → Clicks or pseudo conversions.
• 📊 Observations → Performance data (cost, conversion rates, history, etc.).
• 🗃 Data Format:
  • date, campaign_name, cost, click, ctv, vtv (view-through conversions, if available).
• ⚙ Environment Configurations:
  • Tp = 20 (future window size & stationary days).
  • change_threshold = 1, budget_granularity = 500.

🔹 Cost Control & Reward Modeling:

• 🏦 Monthly Budget Setting → Based on actual total costs consumed.
• 📈 Reward Function → Modeled as a function of cost vs. clicks or pseudo conversions.
• ⚠ Cost Control → Uses Google Ads-style budgeting rules.

---

🔄 Running Experiments

🚀 Running the TUCBMAE Algorithm

python main.py

🏁 Running Baseline Comparisons

python eval_baseline.py

---

License

This project is licensed under the Attribution-NonCommercial 4.0 International License.

Contributing

Contributions are welcome! Please fork the repository and submit a pull request for any improvements or bug fixes

Cite

If you use or reference ABA or the provided dataset, please cite us with the following BibTeX entry:

@article{gangopadhyay2025adaptive,
  title={Adaptive Budget Optimization for Multichannel Advertising Using Combinatorial Bandits},
  author={Gangopadhyay, Briti and Wang, Zhao and Chiappa, Alberto Silvio and Takamatsu, Shingo},
  journal={arXiv preprint arXiv:2502.02920},
  year={2025}
}


📜 **For more details, refer to our research paper or reach out to briti.gangopadhyay@sony.com for assistance!** ✉️