Project-related Links:

LiNGAM-SPP: Causal Discovery Through Shortest Paths

This project explores a novel approach to causal discovery, focusing on reformulating the Linear Non-Gaussian Acyclic Model (LiNGAM) as a shortest path problem, which we refer to as LiNGAM-SPP. This formulation leverages pathfinding algorithms to find the causal ordering of variables even in the presence of unmeasured confounders, significantly enhancing LiNGAM’s accuracy and interpretability. By integrating Pairwise Likelihood Ratios (PLR) instead of traditional mutual information measures, our method improves reliability in causal inference without requiring parameter tuning (vs. existing methods).

The project extends LiNGAM-SPP to include:

• Using pairwise likelihood ratios instead of kNN-based mutual information, removing the need for fine-tuning and improving performance.
• Node-skipping strategies for known relative ordering among nodes.
• Predictive modeling of causal graph properties, including confounder detection, graph sparseness, and LiNGAM-SPP’s performance under different conditions.