A causal discovery algorithm based on the prior selection of leaf nodes.

Area of Science:

• Causal inference and network analysis
• Machine learning and artificial intelligence

Background:

• Linear Non-Gaussian Acyclic Models (LiNGAM) are prevalent for causal network discovery.
• Existing LiNGAM methods face computational complexity and accuracy issues with high-dimensional data or small sample sizes, often due to root node selection.
• These limitations hinder effective causal discovery in complex datasets.

Purpose of the Study:

• To propose a novel causal discovery algorithm, the GPL (Giving Priority to Leaf-nodes) algorithm.
• To address the computational and accuracy limitations of existing LiNGAM methods.
• To provide a more efficient and accurate approach for causal network discovery, particularly for challenging data conditions.

Main Methods:

• Introduced the GPL algorithm, which prioritizes leaf nodes instead of root nodes for causal ordering estimation.
• Leveraged the property that leaf nodes do not influence other nodes to enable direct, bottom-up causal ordering estimation.
• Developed theoretical proofs for the algorithm's feasibility and superiority based on leaf node characteristics.

Main Results:

• The GPL algorithm demonstrated superior performance over state-of-the-art methods in both computational complexity and accuracy.
• Experimental results on synthetic and real-world data confirmed GPL's effectiveness, especially for high-dimensional data (up to 200 variables) and small sample sizes (down to 100 samples for 70 dimensions).
• GPL avoids complex data updating processes inherent in root-node prioritized methods.

Conclusions:

• The GPL algorithm offers a significant advancement in causal discovery, overcoming key limitations of traditional LiNGAM approaches.
• Its leaf-node prioritization strategy provides a more computationally efficient and accurate method for inferring causal networks.
• GPL is particularly advantageous for analyzing complex, high-dimensional datasets with limited sample sizes.