Abstract
Provides the SELF criteria to learn causal structure. Please cite "Ruichu Cai, Jie Qiao, Zhenjie Zhang, Zhifeng Hao. SELF: Structural Equational Embedded Likelihood Framework for Causal Discovery. AAAI. 2018."
Introduction
Causal discovery represents a fundamental challenge in data science, aiming to move beyond statistical correlation to identify the underlying mechanisms of a system. Traditionally, this field has been divided into constraint-based methods, which rely on conditional independence tests, and score-based methods, which search for the optimal graph structure using criteria like the Bayesian Information Criterion (BIC).
However, real-world data often exhibit complex, non-linear relationships and non-Gaussian noise, which can undermine traditional approaches. The SELF (Structural Equation Embedded Likelihood Framework) package addresses these challenges by providing a flexible, likelihood-based framework that can accommodate both linear and complex non-linear causal relationships without requiring a predefined functional form.
Methodology
The SELF framework treats causal discovery as a search for a Directed Acyclic Graph (DAG) that maximizes a likelihood-based score. This process is primarily driven by two components: a search engine and a flexible regression framework.
Fast Hill-Climbing Search
The primary search mechanism is the Fast Hill-Climbing (fhc) algorithm. Starting from an initial graph (typically empty), it iteratively performs edge additions, removals, or reversals to improve the model's score. The algorithm includes parameters to control convergence speed and save intermediate progress.
Flexible Regression Boosters
A key innovation of SELF is its ability to use different "boosters" for modeling structural equations. While standard linear models (lm) are supported, the integration of XGBoost (gbtree) allows the framework to learn complex non-linear relationships directly from the data. This non-parametric approach enables the discovery of interactions that traditional structural equation models might miss.
Users can evaluate candidate graphs using various score types, including BIC, AIC, or log-likelihood. Additionally, the package provides a baseline comparison through mmpcAnm, which combines the Max-Min Parents and Children (MMPC) algorithm with Additive Noise Models (ANM).
Experimental Framework
To facilitate the evaluation of causal discovery algorithms, the SELF package includes tools for generating synthetic benchmark data:
randomGraph: Generates random DAGs with specified node dimensions and average indegrees.synthetic_data_linear: Produces data with linear effects and super-Gaussian noise.synthetic_data_nonlinear: Generates highly complex data using polynomial and trigonometric functions to test the limits of non-linear discovery.
Performance is typically measured using the indicators function, which calculates standard metrics such as the F1-score, Precision, and Recall.
Discussion and Conclusion
The SELF framework offers a powerful approach to causal discovery by unifying linear and non-linear modeling under a single likelihood-based criterion. By leveraging gradient boosting, it provides a data-driven way to capture complex causal mechanisms.
However, users should be aware of certain limitations. As a greedy search algorithm, Fast Hill-Climbing is susceptible to local optima. Furthermore, the computational cost of using non-linear boosters like XGBoost scales with the complexity of the graph. Finally, the framework generally assumes causal sufficiency, meaning it may not account for unobserved confounding variables. Despite these considerations, SELF remains a versatile tool for researchers seeking to uncover the causal structure of their data within the R ecosystem.
Summary updated