Orthogonal Structure Search for Efficient Causal Discovery from Observational Data

TL;DR

This paper introduces a scalable causal discovery method that works with observational data alone, providing theoretical guarantees and handling nonlinear relationships, outperforming existing approaches in large graph scenarios.

Contribution

The proposed algorithm is the first to efficiently infer causal structure from observational data with theoretical guarantees, even in partially nonlinear settings.

Findings

Requires only one estimation per variable

Demonstrates significant improvements over existing methods on large graphs

Handles nonlinear relationships in causal discovery

Abstract

The problem of inferring the direct causal parents of a response variable among a large set of explanatory variables is of high practical importance in many disciplines. Recent work exploits stability of regression coefficients or invariance properties of models across different experimental conditions for reconstructing the full causal graph. These approaches generally do not scale well with the number of the explanatory variables and are difficult to extend to nonlinear relationships. Contrary to existing work, we propose an approach which even works for observational data alone, while still offering theoretical guarantees including the case of partially nonlinear relationships. Our algorithm requires only one estimation for each variable and in our experiments we apply our causal discovery algorithm even to large graphs, demonstrating significant improvements compared to well established approaches.