📊 CUVET-Policy Dataset

From the paper: "CUVET: A Partitioning Approach for Continuous Treatment Assignment At Scale" (under review process at a conference)

📌 Overview

The CUVET-Policy dataset was collected by an online advertising platform conducting a two-week A/B test with five different treatments. The dataset has been properly anonymized to ensure privacy while maintaining its utility for research.

🔹 Treatment Parameters

The dataset includes randomly assigned treatment parameters: {0.8, 0.9, 1, 1.1, 1.2}, where 1 corresponds to the reference treatment.
The goal is to learn a policy that assigns a continuous treatment policy to users, generating more value in expectation than the reference treatment while considering cost constraints.

---

🔐 Privacy & Anonymization

To ensure privacy and business confidentiality, we applied several preprocessing steps:

• The original dataset does not include any data that directly identifies a user, such as names, postal addresses, or email addresses in plain text, so the original data is pseudonymized.
• We manually selected some continuous user features coming from their online interactions, added noise, and standardized them, making the re-identification impossible as the original scale is not provided.
• Value and cost features were scaled and perturbed to maintain confidentiality.

---

📂 Dataset Details

• Size: 86.7M rows (each row represents a user).
• Features:
  • Continuous user features: f0, f1, f2, f3, f4
  • Labels: value and cost
  • Treatment variable: policy
• Splitting strategy:
  • Data is randomly split into train and test sets.
  • The train set contains an additional binary column (is_val) to separate the validation split used in the paper experiments.

---

🔬 Intended Use

The dataset is released under the CC-BY-NC-SA 4.0 license, allowing sharing and adaptation under Attribution, NonCommercial, and ShareAlike conditions.

While our paper focuses on optimal treatment assignment, the dataset can also be used for:
✅ Predict-and-Optimize: Learning models for decision-making under constraints.
✅ Variance-aware methods: Optimizing outcomes in the presence of uncertainty.
✅ Causal Methods: Improving prediction algorithms using noisy experimental data.

---

📢 Citation & Contact

If you use this dataset in your research, please cite our paper:
📄 "CUVET: A Partitioning Approach for Continuous Treatment Assignment At Scale"