因果发现算法研究综述

A Survey on Causal Discovery Algorithms

摘要: 因果发现旨在从观测数据或实验数据中自动化地辨识变量之间的因果关联结构，是因果推理研究领域的核心议题之一。本文系统性地梳理了因果发现的理论根基、算法分类体系及其演进历程。首先，阐明因果发现问题的基本形式化框架，涵盖结构因果模型、有向无环图、因果马尔可夫条件与忠实性假设等核心理论要素；其次，依据方法论范式将现有算法划分为基于约束的方法、基于评分的方法、基于函数因果模型的方法、混合策略方法以及基于深度学习的新兴方法五大类别，逐一剖析其理论动机、代表性算法、技术优势与固有局限；最后，对该领域面临的开放性挑战与未来发展趋向进行展望。

Abstract: Causal discovery aims to automatically identify the underlying causal relationships among variables from observational or experimental data, and has become a central topic in the field of causal inference. This paper provides a systematic review of the theoretical foundations, algorithmic taxonomy, and evolutionary development of causal discovery methods. First, the fundamental formal framework of the causal discovery problem is introduced, including core theoretical components such as structural causal models, directed acyclic graphs, the causal Markov condition, and the faithfulness assumption. Second, according to methodological paradigms, existing algorithms are categorized into five major classes: constraint-based methods, score-based methods, functional causal model-based methods, hybrid approaches, and emerging deep learning-based methods. For each category, the theoretical motivation, representative algorithms, technical advantages, and inherent limitations are systematically analyzed. Finally, the paper discusses the open challenges faced by the field and outlines potential directions for future research.