gCastle Tutorial

A hands-on tutorial for learning causal structure discovery with gCastle, Huawei’s comprehensive causal discovery toolchain. Learn to identify causal relationships in your data using multiple algorithms and evaluation metrics.

Quick Start¶

From the root of the repository:

cd tutorials/gCastle
./docker_build.sh
./docker_jupyter.sh

Then open your browser to http://localhost:8888 and work through the notebooks in order:

1. gCastle.API.ipynb (20 minutes)

   • Learn the core gCastle APIs
   • Generate synthetic causal data
   • Run constraint-based (PC), score-based (GES), and gradient-based (NOTEARS) algorithms
   • Evaluate results with standard metrics (F1, SHD, TPR, FDR)

2. gCastle.example.ipynb (25 minutes)

   • Complete application: discovering causal relationships in economic data
   • Compare algorithm performance on realistic data
   • Interpret and visualize learned causal structures

Key Concepts¶

gCastle provides tools for causal discovery—learning the causal structure of a system from observational data:

• Constraint-based methods (e.g., PC): Use independence tests to discover structure
• Score-based methods (e.g., GES): Optimize over possible DAG structures
• Gradient-based methods (e.g., NOTEARS): Use continuous optimization with acyclicity constraints

Each approach has different computational profiles and assumptions, making them suitable for different data scenarios.