DeDL Framework Toolkit

This repository contains a reusable experiment framework derived from the original synthetic experiment notebook for the paper: Zikun Ye, Zhiqi Zhang, Dennis J. Zhang, Heng Zhang, Renyu Zhang (2025) Deep Learning-Based Causal Inference for Large-Scale Combinatorial Experiments: Theory and Empirical Evidence. Management Science 0(0).

The notebook is preserved for reference, and the reusable Python modules provide a command-line workflow for running new DeDL experiments on synthetic or real data.

Introduction to DeDL Framework

The core logic of the toolkit follows the design approach proposed in Section 3 of the paper, which combines a structured deep neural network (DNN) with Double Machine Learning (DML): first, a structured DNN approximates the nuisance functions in the data-generating process (DGP), and then influence functions are used to correct biases in the predictions, enabling causal inference for unobserved combinations.

Installation

1. Clone this repository:

   clone https://github.com/ysd1123/DeDL_Replication.git
   cd DeDL_Replication

2. (Optional) Create and activate a virtual environment.

   uv venv .venv
   source .venv/bin/activate

3. Install dependencies:

   uv pip install -r requirements.txt

   or if you don't use a virtual environment:

   pip install -r requirements.txt

Quick start

Run the provided synthetic example:

cd ReplicationCodes
python Validation_of_DeDL.py

Results and Visualizations

see visualization/ for scripts to generate figures from the given result data.