VALOR: Value-Aware Learning for Optimized Revenue

VALOR is a research framework for uplift modeling, specifically designed for revenue optimization in B2B and e-commerce scenarios. It leverages Zero-Inflated Log-Normal (ZILN) distributions, Treatment-Gated Interactions (GTI), and Value-Weighted Ranking losses to accurately estimate Individual Treatment Effects (ITE) on revenue.

Repository Structure

VALOR/
├── README.md               # Project documentation
├── requirements.txt        # Python dependencies
├── main.py                 # CLI entry point for experiments
├── config.py               # Hyperparameters and model mappings
├── data/
│   ├── __init__.py
│   ├── loaders.py          # Data preprocessing and DataLoader utilities
│   └── generators.py       # Synthetic data generation logic (ZILN-compatible)
├── models/
│   ├── __init__.py
│   ├── layers.py           # Reusable components (GTI, ZILNHead, MoE)
│   ├── baselines.py        # Baseline models (S/T-Learner, TARNet, DragonNet, etc.)
│   ├── valor.py            # VALOR-specific architectures
│   └── trees.py            # Robust ZILN Forest implementation
├── training/
│   ├── __init__.py
│   ├── losses.py           # ZILN, Focal, and Ranking loss functions
│   └── trainers.py         # Training and validation loops
└── utils/
    ├── __init__.py
    └── metrics.py          # Uplift metrics (AUUC, Qini, Lift@k, Kendall)

Installation

pip install -r requirements.txt

Usage

To run an experiment with the VALOR-enhanced TARNet model on synthetic data:

python main.py --model valor_tarnet --dataset synthetic --runs 5

Supported Models

• Baselines: tlearner, tarnet, dragonnet, cfrnet, cfrnet_wass, unite, euen, causal_forest
• RERUM: rerum (Robust Evaluation and Ranking for Uplift Modeling)
• VALOR Variants: valor_tarnet, valor_dragonnet, valor_unite, valor_euen
• Tree Models: ziln_forest

Key Features

• Focal-ZILN Loss: Specifically designed to handle zero-inflated continuous outcomes like revenue.
• Treatment-Gated Interaction (GTI): A mechanism to learn feature interactions that are sensitive to treatment assignment.
• Value-Weighted Ranking Loss: Optimizes the ranking of users by their expected revenue uplift.
• Bayesian Smoothing: Used in ZILN Forest to handle sparse data in tree leaves.

Citation

If you use this code in your research, please cite the VALOR paper:

@article{valor2026,
  title={VALOR: Value-Aware Learning for Optimized Revenue},
  author={...},
  journal={...},
  year={2026}
}