Reproducibility

InsideForest relies on randomness during tree training and region sampling. Follow these guidelines to produce repeatable experiments.

Fix random seeds

• Set random_state in the underlying forest via rf_params.
• When using NumPy or pandas pipelines, seed them explicitly: np.random.seed(42).
• Use PYTHONHASHSEED=0 when running scripts to keep hashing consistent.

Log environment details

• Capture the output of pip freeze or conda env export in each experiment folder.
• Document the dataset version, preprocessing steps, and any filtering applied before training.
• Record hardware characteristics (CPU cores, RAM) when benchmarking performance.

Persist models and metadata

• Use save / load on canonical clusterers to archive fitted models.
• Store raw_regions_, regions_, region_metrics_, labels_, and the output of region_quality_report.
• Use version control (Git) to track notebooks and configuration files.

Automate checks

• Create smoke tests that run fit and predict on a reduced dataset.
• Assert coverage, AMI, purity, region count, and assignment stability thresholds; include -1.
• Schedule periodic runs with tools such as GitHub Actions to validate dependencies.

Share results

Combine these practices with the logging templates from Experiments & Benchmarks so collaborators can reproduce findings with minimal setup.