Mapping and Advancing the Scalability-Accuracy Frontier of Nonlinear Causal Discovery
SPADE makes combinatorial nonlinear causal discovery run at much larger practical scales without giving up structural accuracy.
The paper compares four major nonlinear causal-discovery families and finds that their runtime and accuracy failures differ by method class. Its proposed spline-based SPADE scheme avoids repeated local scoring by compiling sufficient statistics once and reusing them during combinatorial search. Under bounded indegree, the Gaussian variant cuts complexity from `O(nd^3)` to `O(nd^2+d^3)`. The authors report 100-variable runs with 160K samples in seconds and 1600-variable runs with 2.5K samples in minutes. ArXiv · AI/CL/LG's note
The paper compares four major nonlinear causal-discovery families and finds that their runtime and accuracy failures differ by method class. Its proposed spline-based SPADE scheme avoids repeated local scoring by compiling sufficient statistics once and reusing them during combinatorial search. Under bounded indegree, the Gaussian variant cuts complexity from `O(nd^3)` to `O(nd^2+d^3)`. The authors report 100-variable runs with 160K samples in seconds and 1600-variable runs with 2.5K samples in minutes. ArXiv · AI/CL/LG's note
score 4