Reproducing Main Results

Follow the steps below to reproduce the main results from our paper.

1. Generate the Dataset

Run data_generator.py to generate the synthetic dataset.

• Configuration: Modify config/data_gen.yaml to change tunable parameters such as the number of incidents, variables, and states.
• Internally, this script calls graph_gen.py, which generates individual graph structures using config/graph_gen.yaml.

2. Learn the Prior

Before running RCG(CPDAG), we need to learn the true CPDAG:

• Use learn_kess_g.py with k = -1 to learn the ground truth CPDAG.
• You may specify multiple values of k to generate different k-essential graphs.
• Use learn_prior.yaml for configuration.

Important parameters in learn_kess_g.py:

• k: Value for k-PC (-1 for CPDAG and any positive value as the paramter k for k-PC).
• ORACLE: Whether to use d-separation (oracle) or data-based CI tests.

3. Run the Experiment

Run compare_rcd.py to execute RCG and other baselines:

• Experiment types:
  • Vary the number of nodes and evaluate top-$l$ accuracy (see Figure 4(a)).
  • Vary the number of anomalous samples and evaluate top-1 accuracy (see Figure 4(b)).

Use the --exp flag to switch between experiment types. Possible values are 1 and 2.

• Baselines: To select or modify the set of baselines, edit the BASELINES list in compare_rcd.py.

4. Plot the Results

Visualize experiment results using plot_exp.py.

• Use the --path argument to provide the output directory generated by compare_rcd.py.
• Use the same --exp value as in the previous step to select the correct plot.

Reproducing Results from the Paper

To facilitate reproducibility, we have included the dataset used in our paper. It is available in the UAI-25-dataset directory.

To reproduce the results shown in the paper, run the following commands:

# Reproduce results for Figure 4(a)
python3 compare_rcd.py --path UAI-25-dataset --exp 1
python3 plot_exp.py --path UAI-25-dataset/exp_results/{PATH} --exp 1

# Reproduce results for Figure 4(b)
python3 compare_rcd.py --path UAI-25-dataset --exp 2
python3 plot_exp.py --path UAI-25-dataset/exp_results/{PATH} --exp 2

Replace {PATH} with the specific subdirectory created during the experiment run.

The final plots used in the paper can be found in UAI-25-dataset/exp_results/nodes and UAI-25-dataset/exp_results/int-samples.

---

Important Files and Folders

File / Folder	Description
graph_gen.py	Generates a single graph (incident) using config/graph_gen.yaml.
data_generator.py	Generates the full dataset by calling graph_gen.py multiple times. Configured via config/data_gen.yaml.
learn_kess_g.py	Learns the k-essential or CPDAG graph from an incident.
para_kpc/	Parallel implementation of k-PC using multiprocessing.
learn_prior.py	Learn the k-essential or CPDAG graph for a dataset.
m_igs.py	Implementation of M-IGS.
rcg.py	Implementation of RCG.
compare_rcd.py	Runs multiple baselines on a given dataset and stores the result.
plot_exp.py	Plots the results generated by compare_rcd.py.

---

Additional Notes

• CI Tests: We use chi-square (chisq) as the default CI test, which is suitable for discrete data.
• Discretization: For real-world datasets (e.g., Sock-shop), discretize continuous data using the BINS parameter.
• Preprocessing: Use the PRE_PROCESS flag to remove irrelevant columns like timestamps or constant-value features. Look for function _select_cols and discretize in base_utils.py for more information.
• Parallelism: Most scripts support parallel execution via THREADING and WORKERS parameters.