cdzoo

cdzoo puts a portfolio of causal-discovery algorithms behind one interface. R has strong individual packages -- pcalg, bnlearn -- but no single harness that wraps R and Python algorithms in one convention, with a scenario grid and fair metrics. cdzoo provides:

• a registry of nineteen discovery methods drawn from pcalg, bnlearn, and Python back-ends (LiNGAM, causal-learn, gCastle, DAGMA) reached through reticulate, of which sixteen are scored in the shipped field benchmark; each returns its estimate in one adjacency convention;
• an evidence-routed recommender, cd_auto(), that inspects a data set, routes it to a method suited to that regime, runs it, and reports its rationale and a confidence label -- a recommender, not a selector, because causal discovery has no run-time ground truth;
• a scenario grid of structural-causal-model regimes with known DAGs, and metrics (skeleton F1, structural Hamming distance, orientation recall) to score recovered graphs.

Every back-end is optional. The harness, the scenarios, the metrics, and the recommender's routing logic all run on a base installation; a method runs only when its back-end is present.

Related packages

• pcalg learns causal structure from observational and interventional data with PC, FCI, RFCI and GIES, and estimates causal effects with IDA and adjustment criteria.
• bnlearn learns the structure and parameters of Bayesian networks with constraint-based, score-based and hybrid algorithms, and runs inference on the fitted network.
• causalDisco learns causal structure from temporally ordered observational data, adds background knowledge to the pcalg, bnlearn and Tetrad algorithms, and compares and evaluates the learned graphs.

cdzoo calls pcalg and bnlearn as back-ends rather than replacing them. It adds Python methods to the same registry, scores every method on scenarios with known graphs, and recommends a method for a given data set with the reasons for the choice.

Installation

# install.packages("pak")
pak::pak("max578/cdzoo")

The R back-ends are installed with install.packages(c("pcalg", "bnlearn")). The Python back-ends are reached through reticulate; install them into the active Python environment (for example lingam, causal-learn, gcastle, dagma).

cdzoo never starts Python just to report availability: the Python back-ends show as installed once reticulate's interpreter is initialised or configured (for example via RETICULATE_PYTHON). Set options(cdzoo.python_init = TRUE) to let the availability check initialise Python eagerly, and call list_methods(refresh = TRUE) after installing a Python module mid-session.

Usage

library(cdzoo)

# what is registered, and which back-ends are installed?
list_methods()

# generate data from a known DAG, then recover it with one method
sim <- cd_scenario("linear_laplace", seed = 1)
res <- cd_discover(sim$data, "icalingam_pcalg")
cd_score(res, sim$truth)

# or let the recommender choose, and explain its choice
rec <- cd_auto(sim$data)
rec@method
rec@rationale

cd_auto() also accepts an evaluable density -- a proxymix mixture or target -- rather than a sample, reading the conditional-independence structure off the mixture in closed form, the case no sampling-based method can serve.

Benchmark evidence

cd_field_benchmark ships a summary of the sixteen methods across nine structural-causal-model scenarios, scored against each scenario's known DAG. It is the evidence behind the recommender's routing, and the worked example in the package vignette.

Contributing

Bug reports and suggestions are welcome through the issue tracker.

Citation

citation("cdzoo")

Licence

MIT. See LICENSE.md.