FLOP Causal Discovery Algorithm

This repository contains a Rust implementation of the FLOP causal discovery algorithm, available for use from Python and R. It is a score-based algorithm for learning equivalence classes of DAGs from observational data, assuming linear relationships between variables.

Installation

In Python, flopsearch can be installed via pip:

pip install flopsearch

In R, flopsearch can be installed directly from GitHub:

install.packages("https://github.com/CausalDisco/flopsearch/releases/download/v0.3.0/flopsearch.tar.gz")

This requires a working installation of the Rust toolchain.

The name of the installed package is flopsearch, and it can be loaded with:

library(flopsearch)

Citing FLOP

If you use FLOP in your scientific work, please cite this paper:

@article{embracing2026,
  author  = {Marcel Wien{\"o}bst and Leonard Henckel and Sebastian Weichwald},
  title   = {{Embracing Discrete Search: A Reasonable Approach to Causal Structure Learning}},
  journal = {International Conference on Learning Representations (ICLR)},
  year    = {2026}
}

How To Run FLOP

In Python, as a simple example, FLOP can be called by

flopsearch.flop(X, 2.0, restarts=50)

with X being the data matrix, 2.0 the BIC penalty parameter and the number of ILS restarts being set to 50.

Similarly, in R, one can call:

flopsearch::flop(X, 2.0, restarts=50)

Instead of the number of restarts, it is also possible to set a timeout in seconds after which the search terminates and returns the best-scoring graph found thus far.

FLOP returns a CPDAG encoded with an adjacency matrix whose entry in row i and column j is 1 in case of a directed edge from the i-th to the j-th variable and 2 in case of an undirected edge between those variables (in case of an undirected edge, the entry in row j and column i is also 2, that is each undirected edge induces two 2's in the matrix).