Abstract:With a large amount of data accumulated in many fields, causal discovery based on observational data is gradually emerging, which is considered to be the basis for realizing strong artificial intelligence. However, the existing main causal discovery methods, including constraint-based methods, structural causal model based methods, and scoring-based methods, cannot find real causal relations accurately and quickly. In this paper, we propose a causal discovery method based on genetic algorithm, which combines structural causal model, scoring method, and genetic search algorithm. The core of our method is to divide the causal relation discovery process into the evaluation phase based on the features of structural causal model and the search phase based on the genetic algorithm. In the evaluation phase, the causal graph is evaluated from three aspects: model deviation, noise independence, and causal graph cyclicity, which effectively ensures the accuracy of causal discovery. In the search phase, an efficient random search is designed based on genetic algorithm, which greatly improves the causal discovery efficiency. This paper implements the corresponding algorithm, namely SCM-GA (Structural Causal Model based Genetic Algorithm), and conducts experiments on several simulated datasets and one widely used real-scene dataset. The experiments compare five classic baseline algorithms, and the results show that SCM-GA has achieved great improvement in accuracy, applicability, and efficiency. Especially on the real scene dataset, SCM-GA achieves better results than the state-of-the-art algorithm, with similar SHD (Structure Hamming Distance) value, 40% higher recall rate, and 83.3% shorter running time.

An Entropy Based Method for Causal Discovery in Linear Acyclic Model.

Local Causal Discovery with Linear non-Gaussian Cyclic Models

Functional Linear Non-Gaussian Acyclic Model for Causal Discovery

A Genetic Algorithm for Causal Discovery Based on Structural Causal Model.

Causal Discovery in Linear Non-Gaussian Acyclic Model With Multiple Latent Confounders

Partial Homoscedasticity in Causal Discovery with Linear Models

Directed Cyclic Graph for Causal Discovery from Multivariate Functional Data

Causal Discovery of Linear Non-Gaussian Causal Models with Unobserved Confounding

Causal Discovery in Linear Structural Causal Models with Deterministic Relations

Causal Graph Discovery from Self and Mutually Exciting Time Series

Induced Covariance for Causal Discovery in Linear Sparse Structures

Nonlinear Causal Discovery in Time Series

Hybrid Top-Down Global Causal Discovery with Local Search for Linear and Nonlinear Additive Noise Models

Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data

Entropy-based Discovery of Summary Causal Graphs in Time Series

Confidence in Causal Discovery with Linear Causal Models

Scalable Variational Causal Discovery Unconstrained by Acyclicity

Causal Layering via Conditional Entropy

Combining Linear Non-Gaussian Acyclic Model with Logistic Regression Model for Estimating Causal Structure from Mixed Continuous and Discrete Data

Causal discovery from sequential data in ALS disease based on entropy criteria

Causal Discovery in Linear Models with Unobserved Variables and Measurement Error