Reproducing the paper experiments¶

The experiments/ folder in the source repo contains scripts that re-run the experiments from the CDANs paper using this library, on the same data files used by the MATLAB reference implementation.

Quick run¶

From the repo root, with the parent MATLAB repo's data folder available:

python -m experiments.ot   --data-dir ../CDANs/OT
python -m experiments.x4_4 --data-dir ../CDANs/x4_4

Each script prints a structured report: lagged edges, contemporaneous edges (directed and undirected), changing modules, and per-step timings.

Available experiments¶

Name	Variables	Time points	MATLAB script
`OT`	12	86	`OT_CDANs.mlx`
`x4_4`	8	347	`x4_4_CDANs.mlx`

The other experiments in the parent repo (x4_6, x6_4, x8.8) follow the same pattern; adding them is a five-line change in experiments/loaders.py (register a new ExperimentSpec) plus a thin runner script.

Parameter mapping (MATLAB → Python)¶

The defaults in experiments/runner.py mirror the MATLAB driver settings:

MATLAB	Python (`CDANs(...)`)
`alpha = 0.05`	`alpha=0.05`
`maxFanIn = 4`	`max_extra_conds=4`
`cond_ind_test = 'indtest_new_t'`	`ci_test="kci"`
`pars.if_GP1 = 1` (small T)	KCI default (fixed-bandwidth kernel)
`pars.if_GP2 = 1`	independent-change uses fixed `0.1`
`pars.width = 0` → MATLAB sets `0.1`	`independent_change_width=0.1`
`Type = 1` (phases 1+2+3)	`use_independent_change=True`

Caveats¶

Step 1 isn't in the MATLAB scripts. The MATLAB drivers feed pre-lagged data into a modified CD-NOD that uses a hard-coded initial graph (nonsta_cd_new.m line 88). The Python library runs the full pipeline, including Step 1 (PCMCI lagged-adjacency discovery), so it discovers its own lagged structure rather than the paper's hand-curated initial graph. Use --tau-max 1 to keep this minimal, or higher values to look further back.
GP-learned bandwidths: partially ported. The MATLAB code uses if_GP1=1 for KCI and if_GP2=1 for the independent-change kernel when T <= 1000. The independent-change side is now ported via independent_change_width="gp", which fits an ARD-RBF Gaussian process via scikit-learn and uses the marginal-likelihood-optimized length scales as bandwidths. The KCI test still uses fixed-bandwidth Gaussian kernels — that's the remaining gap. Results should be qualitatively similar to the MATLAB reference but won't be bit-exact.
Output comparison is left to the caller. The MATLAB experiments save output.mat with the recovered adjacency matrices; comparing them programmatically would require scipy.io.loadmat.

Adding a new experiment¶

Open experiments/loaders.py and register a new ExperimentSpec constant with the experiment name, the CSV filenames in the order used by the original MATLAB script, and the expected time-series length.
Add it to the EXPERIMENTS dict in the same file.
Create a thin runner script next to ot.py that imports your spec and calls cli(your_spec).

That's it — runner.py does everything else.