ETIA: A Comprehensive Automated Causal Discovery Library

Find the complete Documentantion https://etia.readthedocs.io/en/latest/index.html#

Library Overview

ETIA (Αιτία (pronounced etía): "cause" in Greek) is a cutting-edge automated causal discovery library that takes causal analysis beyond traditional methods. It is designed to tackle complex, real-world problems by automating the entire causal discovery process, offering a combination of feature selection, causal structure learning, and causal reasoning validation that is unmatched in other libraries.

ETIA provides:

• Dimensionality Reduction through its Automated Feature Selection (AFS) module, which ensures only the most relevant variables are used, even in high-dimensional datasets.
• Causal Structure Learning via the Causal Learning (CL) module, automating the discovery of the causal graph with the best possible fit to the data.
• Causal Reasoning Validation (CRV) offers tools for computing confidence in discovered relationships and visualizing causal paths.

Unlike existing libraries, ETIA does not simply offer isolated algorithms—it provides a fully automated pipeline that optimizes and customizes each step of the causal discovery process, ensuring the results are robust, interpretable, and reliable for both researchers and industrial practitioners.

Why ETIA is Unique

ETIA goes beyond other causal discovery libraries by offering:

• End-to-End Automation: ETIA fully automates the discovery process, combining various algorithms and approaches to find the best configuration for a given dataset. This level of automation is rare in other libraries, which often leave algorithm selection and tuning to the user.
• Out-of-Sample Causal Tuning: A method developed specifically for ETIA, Out-of-Sample Causal Tuning ensures the best causal graph is selected without the need for manually tuned parameters. This is essential when working in unsupervised environments where traditional estimation methods, like K-fold cross-validation, fail.
• Confidence Estimation and Visualization: ETIA not only discovers causal graphs but also evaluates the confidence of each relationship within the graph. Bootstrapped confidence metrics and visualization tools help users understand which findings are most reliable.
• Dimensionality Reduction with Causal Insight: The AFS module uses sophisticated techniques to reduce the dimensionality of datasets without sacrificing causal accuracy, a critical need when dealing with high-dimensional, real-world data.
• Handling Latent Variables: Unlike many causal discovery tools, ETIA can identify causal relationships even in datasets with hidden confounders through algorithms like FCI and GFCI.

Core Features

1. Automated Feature Selection (AFS)

AFS goes beyond standard feature selection by targeting the Markov Boundary of the outcome of interest. This approach ensures that you work only with the most causally relevant variables, preventing the noise and redundancy that plague other methods.

Algorithm	Description	Data Type
FBED	Forward-Backward selection with Early Dropping	Mixed
SES	Statistical Equivalence Selection	Mixed

Why It Matters: In high-dimensional datasets, choosing the right features is critical. AFS uses state-of-the-art techniques to find causally relevant features, ensuring that the subsequent causal analysis is accurate and manageable, even in datasets with hundreds of variables.

2. Causal Learning (CL)

The CL module identifies causal relationships using a variety of algorithms that are automatically optimized to your data. ETIA's causal tuning mechanism guarantees the selection of the best possible causal structure without the need for user intervention.	Algorithm	Latent Variables Supported	Tests/Scores Used	Data Type
PC Algorithm	✕	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
CPC	✕	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
FGES	✓	SEM BIC Score, BDeu, Discrete BIC, CG BIC, DG BIC	Continuous, Mixed, Categorical
FCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
FCI-Max	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
RFCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
GFCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
CFCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical
sVAR-FCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square	Continuous, Mixed, Categorical (Time Series)
svargFCI	✓	FisherZ, CG LRT, DG LRT, Chi-square, G-square, SEM BIC Score, BDeu, Discrete BIC, CG BIC, DG BIC	Continuous, Mixed, Categorical (Time Series)
PCMCI	✕	ParCor, RobustParCor, GPDC, CMIknn, ParCorrWLS, Gsquared, CMIsymb, RegressionCI	Continuous, Mixed, Categorical (Time Series)
PCMCI+	✕	ParCor, RobustParCor, GPDC, CMIknn, ParCorrWLS, Gsquared, CMIsymb, RegressionCI	Continuous, Mixed, Categorical (Time Series)
LPCMCI	✓	ParCor, RobustParCor, GPDC, CMIknn, ParCorrWLS, Gsquared, CMIsymb, RegressionCI	Continuous, Mixed, Categorical (Time Series)
SAM	✕	Learning Rate (lr), Decay Learning Rate (dlr), Regularization (lambda1, lambda2), Hidden Neurons (nh, dnh), Training Epochs (train_epochs), Testing Epochs (test_epochs), Batch Size (batch_size), Loss Type (losstype)	Continuous, Mixed
NOTEARS	✕	Max Iterations (max_iter), Tolerance (h_tol), Threshold (threshold)	Continuous, Mixed, Categorical

Key Details:

• Latent Variables Supported:

  • ✓: Supports latent (unobserved) variables.
  • ✕: Does not support latent variables (causal sufficiency assumed).

• Tests/Scores Used:

  • Conditional Independence Tests (ci_test): Methods like FisherZ, CG LRT, DG LRT, Chi-square, G-square.
  • Scores (score): Metrics like SEM BIC Score, BDeu, Discrete BIC, CG BIC, DG BIC.
  • Additional Parameters: Algorithms like SAM and NOTEARS have specific parameters relevant to their optimization and learning processes.

• Data Type:

  • Continuous: Numeric data without discrete categories.
  • Mixed: Combination of continuous and categorical data.
  • Categorical: Data with discrete categories.
  • Time Series: Data that includes temporal dependencies.

Notes:

• Assumptions:
  • Causal Sufficiency: If set to False, the algorithm accounts for potential latent variables.
  • Assume Faithfulness: Indicates whether the algorithm assumes the faithfulness condition holds, impacting its ability to recover the true causal graph.

Why It Matters: Traditional causal discovery libraries expect users to manually choose an algorithm. ETIA’s automated pipeline selects the best algorithm for your dataset, saving time and reducing the risk of suboptimal results. Additionally, it can handle datasets with latent variables, which most other systems cannot do.

3. Causal Reasoning Validation (CRV)

CRV provides advanced tools to evaluate the discovered causal graph, offering confidence estimates and comprehensive visualizations. It can answer specific causal queries, making it an invaluable tool for decision-makers and researchers alike.

Functionality	Description
Visualization	Visualize graphs and causal relations using Cytoscape.
Adjustment Sets	Identify adjustment sets needed for estimating causal effects.
Confidence Calculations	Assess confidence in discovered causal relationships through bootstrapping methods.
Causal Queries	Answer user-defined causal queries, including directed, bidirected, and potentially directed paths between variables.

Why It Matters: The ability to compute and visualize confidence in causal relationships sets ETIA apart from other libraries. Users can trust that the discovered causal graph is not just a hypothesis but a statistically backed structure with clearly defined confidence levels.

Installation

You can install ETIA directly from PyPi using pip:

pip install etia

Alternatively, clone the repository and install the dependencies:

git clone <repository-url>
cd library directory
pip install -r requirements.txt
make all

Prerequisites

Before installing ETIA, ensure that you have the following dependencies:

• Python 3.8+
• Java 17 (required for Tetrad algorithms in the Causal Learning module)
• R 4.4+ (required for some feature selection algorithms in the AFS module)
• Cytoscape (required for the visualization)
• MxM package in R (required for AFS, more information on that follows)
• daggity package in R (required for CRV.adjustment_set, more information on that follows)

You can download and install these dependencies from their official websites:

• Python
• Java
• R
• Cytoscape

ETIA Installation

Installing via PyPi (Upcoming)

Once ETIA is available on PyPi, you will be able to install it directly using pip:

 pip install etia

Installing from Source

To install ETIA from the source code, follow the steps below:

1. Clone the repository:

 git clone <repository-url>
 cd etia

2. Install the required dependencies:

 pip install -r requirements.txt

3. Compile the necessary components:

make all

Java and R Configuration

For using Tetrad algorithms and certain feature selection algorithms, ensure that Java and R are correctly installed.

Setting up Java:

1. Install Java 17 from the official website.
2. Ensure that the JAVA_HOME environment variable is set:

 export JAVA_HOME=/path/to/java   

Setting up R:

1. Install R 4.4+ from the official website.
2. Make sure that R is in your system’s PATH:

R --version

3. Install MxM package (this part may take a while), and the daggity package. MxM package is necessary for AFS while daggity is only used in CRV to find the adjustment sets. Note: Depending on the OS you may need to install CMake and GSL (GNU Scientific Library)

Rscipt --vanilla "install.packages("MXM", repos = "http://cran.us.r-project.org")"
Rscipt --vanilla "install.packages("daggity", repos = "http://cran.us.r-project.org")"

Usage

Once installed, ETIA can be used by importing its modules. Here is a simple example for feature selection and causal discovery:

from ETIA.AFS import AFS
from ETIA.CausalLearning import CausalLearner

# Feature selection
afs = AFS()
results = afs.select_features(dataset_file_path, targets)
reduced_datset = results['reduced_data']

# Causal discovery
cl = CausalLearner(reduced_datset)
results = cl.learn_model()

Testing (under development)

You can run the test suite using:

pytest tests/

Make sure that all dependencies, including Java and R, are correctly installed before running tests.

Contributing

Contributions are welcome! To contribute:

1. Fork the repository.
2. Create a new branch (git checkout -b feature-branch).
3. Commit your changes (git commit -am 'Add new feature').
4. Push to the branch (git push origin feature-branch).
5. Create a Pull Request.

License

This project is licensed under the MIT License. See the LICENSE file for details.