This is an early release of ππππππ π₯ and feedback is very welcome!
If you publish research using ππππππ, please cite our article
@article{henckel2024adjustment,
title = {{Adjustment Identification Distance: A gadjid for Causal Structure Learning}},
author = {Leonard Henckel and Theo WΓΌrtzen and Sebastian Weichwald},
journal = {{arXiv preprint arXiv:2402.08616}},
year = {2024},
doi = {10.48550/arXiv.2402.08616},
}
Just pip install gadjid
to install the latest release of ππππππ from PyPI
and run python -c "import gadjid; help(gadjid)"
to get started.
Pip tries to find a matching wheel and install that.
Since we offer precompiled wheels
for most common operating systems, python versions, and CPU architectures,
the installation will usually be quick.
If there is no matching wheel
(or when explicitly installing from source via
pip install gadjid --no-binary gadjid
),
pip will download the source distribution and compile a wheel for the current platform,
which requires the rust toolchain to be installed.
The current development version can be compiled and installed via
pip install "git+https://github.com/CausalDisco/gadjid.git"
or by cloning this repository and calling either
maturin develop --manifest-path ./gadjid_python/Cargo.toml
(unoptimized dev compile)
or
maturin develop --manifest-path ./gadjid_python/Cargo.toml --release
(optimized release compile).
import gadjid
from gadjid import example, ancestor_aid, oset_aid, parent_aid, shd
import numpy as np
help(gadjid)
example.run_parent_aid()
Gtrue = np.array([
[0, 1, 1, 1, 1],
[0, 0, 1, 1, 1],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]
], dtype=np.int8)
Gguess = np.array([
[0, 0, 1, 1, 1],
[1, 0, 1, 1, 1],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]
], dtype=np.int8)
print(ancestor_aid(Gtrue, Gguess, edge_direction="from row to column"))
print(shd(Gtrue, Gguess))
ππππππ is implemented in Rust and can conveniently be called from Python via our Python wrapper (implemented using maturin and PyO3).
Evaluating graphs learned by causal discovery algorithms is difficult: The number of edges that differ between two graphs does not reflect how the graphs differ with respect to the identifying formulas they suggest for causal effects. We introduce a framework for developing causal distances between graphs which includes the structural intervention distance for directed acyclic graphs as a special case. We use this framework to develop improved adjustment-based distances as well as extensions to completed partially directed acyclic graphs and causal orders. We develop polynomial-time reachability algorithms to compute the distances efficiently. In our package ππππππ, we provide implementations of our distances; they are orders of magnitude faster than the structural intervention distance and thereby provide a success metric for causal discovery that scales to graph sizes that were previously prohibitive.
ππππππ uses rayon for parallelism
using, per default, as many threads as there are physical CPU cores.
The number of threads to use can be set via the environment variable RAYON_NUM_THREADS
.
We recommend to do so and to set the number of threads manually,
not least to be explicit and to avoid the small runtime overhead for determining the number of physical CPU cores.
- Feedback is very welcome! Just open an issue on here.
- We are working on making ππππππ available also for R.
- The code is well documented. We plan on making a user and developer documentation available. π
ancestor_aid(Gtrue, Gguess, edge_direction)
oset_aid(Gtrue, Gguess, edge_direction)
parent_aid(Gtrue, Gguess, edge_direction)
- for convenience, the following distances are implemented, too
shd(Gtrue, Gguess)
sid(Gtrue, Gguess, edge_direction)
β only for DAGs!
where Gtrue
and Gguess
are adjacency matrices of a DAG or CPDAG
and edge_direction
determines whether a 1
at r-th row and c-th column of an adjacency matrix
codes the edge r β c
(edge_direction="from row to column"
) or c β r
(edge_direction="from column to row"
).
The functions are not symmetric in their inputs:
To calculate a distance,
identifying formulas for causal effects are inferred in the graph Gguess
and verified against the graph Gtrue
.
Distances return a tuple (normalised_distance, mistake_count)
of the fraction of causal effects inferred in Gguess that are wrong relative to Gtrue, normalised_distance
,
and the number of wrongly inferred causal effects, mistake_count
.
There are normalised_distance = mistake_count / p(p-1)
.
You may also calculate the SID between DAGs via parent_aid(DAGtrue, DAGguess, edge_direction)
,
but we recommend ancestor_aid
and oset_aid
and for CPDAG inputs the parent_aid
does not coincide with the SID
(see also our accompanying article).
If edge_direction="from row to column"
, then
a 1
in row r
and column c
codes a directed edge r β c
;
if edge_direction="from column to row"
, then
a 1
in row r
and column c
codes a directed edge c β r
;
for either setting of edge_direction
,
a 2
in row r
and column c
codes an undirected edge r β c
(an additional 2
in row c
and column r
is ignored;
one of the two entries is sufficient to code an undirected edge).
An adjacency matrix for a DAG may only contain 0s and 1s. An adjacency matrix for a CPDAG may only contain 0s, 1s and 2s. DAG and CPDAG inputs are validated for acyclicity. However, for CPDAG inputs, the user needs to ensure the adjacency matrix indeed codes a valid CPDAG (instead of just a PDAG).
Experiments run on a laptop with 8 GB RAM and 4-core i5-8365U processor.
Here, for a graph with
Maximum graph size feasible within 1 minute
Method | sparse | dense |
---|---|---|
Parent-AID | 13005 | 960 |
Ancestor-AID | 8200 | 932 |
Oset-AID | 546 | 250 |
SID in R | 255 | 239 |
Results obtained with ππππππ v0.0.1 using the Python interface and the SID R package v1.1 from CRAN.
Average runtime
Method | x-sparse ( |
sparse ( |
dense ( |
---|---|---|---|
Parent-AID | 6.3 ms | 22.8 ms | 189 ms |
Ancestor-AID | 2.7 ms | 38.7 ms | 226 ms |
Oset-AID | 3.2 ms | 4.69 s | 47.3 s |
SID in R | ~1β2 h | ~60 s | ~60 s |
Results obtained with ππππππ v0.0.1 using the Python interface and the SID R package v1.1 from CRAN.
- .github/workflows/ β github actions for linting/testing/packaging
- gadjid/ β Rust core package, which implements
a graph memory layout purposefully designed for fast memory access in reachability algorithms,
new reachability algorithms to check the validity of an adjustment set,
and all DAG/CPDAG distances discussed in the accompanying article
- gadjid/src/snapshots β
ππππππ is extensively tested (tests at bottom of each
/gadjid/src/**.rs
file) and validated against the R implementation of the SID on pairs of DAG inputs (cf. bottom ofparent_aid.rs
); this folder holds insta snapshots for testing graph and reachability algorithms and all distances against (cf. gadjid/src/lib.rs)
- gadjid/src/snapshots β
ππππππ is extensively tested (tests at bottom of each
- gadjid_python/ β
python wrapper that accepts numpy and scipy int8 matrices as graph adjacency matrices
- gadjid_python/tests/ β runs tests of and via the python ππππππ wrapper:
- tests the loading of numpy arrays and views as well as scipy sparse csr/csc matrices
- tests
parent_aid
against the R implementation of the SID on pairs of DAG inputs; since in the special case of DAG inputs the Parent-AID coincides with the SID, this end-to-end tests the check for validity of adjustment sets implemented via new reachability algorithms
- gadjid_python/tests/ β runs tests of and via the python ππππππ wrapper:
- gadjid_r/ β placeholder for the R wrapper to come!
- testgraphs/ β testgraphs in .mtx files (Matrix Market Exchange Format), csv files with the SHD/SID between the testgraphs to test against, checksums
ππππππ is available in source code form at https://github.com/CausalDisco/gadjid.
This Source Code Form is subject to the terms of the Mozilla Public License, v. 2.0. If a copy of the MPL was not distributed with this file, You can obtain one at https://mozilla.org/MPL/2.0/.
See also the MPL-2.0 FAQ.