Mdls35/causal_synthetic_scenarios

Causal Synthetic Scenarios

This directory provides tools for generating synthetic datasets to benchmark causal discovery and inference methods, as implemented in the methodology of the KNOSYS-D-25-17892 paper. The source code for these generators is available in the CausalMorph repository.

---

Naming Convention of Generated Datasets

Each synthetic dataset is saved as two files: one for the adjacency matrix and one for the observed data matrix. The filenames systematically encode all key dataset generation parameters, allowing for easy identification and reproducibility. The naming is produced exactly as in generate_synthetic.py, and follows this structure:

model_r-{rep}_p-{num_vars}_pconn-{pconn}_{dist_name}_deviat-{deviation}_n-{nsamples}_mode-{mode}[_nl-{nonlinearity}]_edges-{n_edges}_dens-{density}_seed-{rep_seed}[_LiNGAM-ideal]-am.csv
model_r-{rep}_p-{num_vars}_pconn-{pconn}_{dist_name}_deviat-{deviation}_n-{nsamples}_mode-{mode}[_nl-{nonlinearity}]_edges-{n_edges}_dens-{density}_seed-{rep_seed}[_LiNGAM-ideal]-dat.csv

where:

• {rep}: repetition/run index for the same parameter set.
• {num_vars}: number of variables (nodes) in the DAG.
• {pconn}: edge probability controlling DAG density.
• {dist_name}: summarises the marginal noise distribution types (e.g., normal-100 means 100% normal).
• {deviation}: additive noise scale (σ for normal, a for uniform etc).
• {nsamples}: number of generated samples (rows).
• {mode}: "linear" or "nonlinear".
• [_nl-{nonlinearity}]: only present if mode is "nonlinear", it reports the degree of nonlinearity (e.g. 0.75).
• {n_edges}: total number of DAG edges.
• {density}: edge density (fraction of possible edges present, to three decimal places).
• {rep_seed}: random seed for full reproducibility of this repetition.
• [_LiNGAM-ideal]: appended only for the "ideal"/control LiNGAM scenario (i.e., mode="linear", normal_ratio=1.0, nonlinearity=0.0).
• The suffix -am.csv means "adjacency matrix" file, and -dat.csv means data matrix file.

Example filename:

model_r-3_p-25_pconn-0.5_normal-100_deviat-0.25_n-5000_mode-linear_edges-31_dens-0.105_seed-99_LiNGAM-ideal-am.csv
model_r-3_p-25_pconn-0.5_normal-100_deviat-0.25_n-5000_mode-linear_edges-31_dens-0.105_seed-99_LiNGAM-ideal-dat.csv

---

Generation Options and Pipeline Overview

The generation pipeline allows exhaustive parameter sweeps over key aspects, including:

• Number of variables: e.g., 5, 25, 40
• Edge probability (pconn): e.g., 0.05, 0.25, 0.5, 0.75
• Noise types: controlled by "normal ratio", each variable gets a noise type ("normal", "uniform", "laplace", "exponential")
• Deviation: controls noise scale; see "Noise Distribution Table" below
• Samples: number of rows
• Seed: for reproducibility; unique to each repetition
• Mode / Nonlinearity: "linear" or "nonlinear", and nonlinearity degree if nonlinear

Trivial settings (e.g., all-normal noise and zero deviation) may be skipped at batch-generation time, and both linear and nonlinear mechanisms are supported.

The filenames allow you to uniquely trace back all these settings for any given file without inspecting its contents.

---

File Contents

Each dataset folder contains:

• *-am.csv: adjacency matrix (shape: p × p; pandas CSV)
• *-dat.csv: causal sample data (shape: n × p; pandas CSV)

---

Example Usage

Generate one dataset (Python):

from causalmorph.data_generation.synthetic_scenarios import causal_graph_synthetic_scenarios

adj_matrix, data = causal_graph_synthetic_scenarios(
    p=25, pconn=0.5, dist=["normal"]*25, deviation=0.5,
    nsamples=500, mode="nonlinear", nonlinearity=0.75, seed=42
)

Batch generation scripts output folders filled with files named as above, automatically.

---

Parameter Reference

Parameter	Type	Description
p	int	Number of variables/nodes.
pconn	float	Edge probability (graph density).
dist	list/str	Per-variable noise types.
deviation	float	Controls noise scale for each variable.
nsamples	int	Number of samples to generate.
seed	int	Random seed for reproducibility.
mode	str	"linear" or "nonlinear" mechanisms.
nonlinearity	float	α for nonlinear mixture (if nonlinear).
rep	int	Repetition/run index.

Other parameters are also encoded as needed. See code for further details.

---

Noise Distribution Table

Distribution	Parameter	Formula	Comment
Normal	σ	ε ~ N(0, σ²)	σ = deviation
Uniform	a	ε ~ U(-a, a)	a = deviation
Laplace	b	ε ~ Laplace(0, b)	b = deviation
Exponential	scale	ε ~ Exp(λ=1/deviation)	scale = deviation

---

Further Notes

• All random seeds, distributions, nonlinearities, and graph structures are encoded in the name or saved to disk for full reproducibility.
• The filename convention enables automated analysis, grouping, and reproducibility without further metadata files.
• For latest usage, utility scripts, and updates, see CausalMorph.

---