Performance Tips

InsideForest includes automated heuristics, but tuning a few parameters can significantly improve runtime and memory usage. Use the following guidelines when working with large datasets.

Leverage auto presets

• auto_fast=True: enables internal shortcuts that reduce the number of evaluated regions and increases sampling efficiency.
• auto_feature_reduce=True: keeps only the most discriminative features per region, lowering dimensionality before clustering.
• n_sample_multiplier and ef_sample_multiplier: control how many samples feed the branch extraction step. Lower values decrease runtime but may miss rare patterns.

Control tree complexity

• Limit max_depth or min_samples_leaf in the underlying forest to avoid overly fragmented branches.
• Use rf_params to pass custom RandomForest settings such as n_estimators, max_features, or bootstrapping.
• Reuse an existing forest when you already have a calibrated model; this skips training time completely.

Batch predictions

• InsideForest estimators accept NumPy arrays, pandas DataFrames, and Spark DataFrames.
• For pandas inputs, call predict, transform, or assign_regions in batches to control membership-matrix memory.
• The region clusterers do not expose predict_proba. Read class distributions from regions_; inspect forest_ separately only when diagnosing the branch generator.

Control the number of regions

• Raise leaf_percentile or min_support to retain fewer, stronger class regions.
• Use max_regions_per_class to bound class-aware output size.
• Keep branch_aggregation="none" unless a validated aggregation strategy meets compression, coverage, AMI, purity, stability, and runtime criteria.

Monitor feature importance

• Inspect feature_importances_ to spot features that dominate rules. Consider removing low-signal columns to simplify the search space.
• Combine auto_feature_reduce with domain knowledge to cap the maximum number of conditions per region.

Persist intermediate artifacts

• Call save on clusterers or regressors to store the forest, regions, and metadata. Reload them with load to resume analysis.
• Export region descriptions to CSV or JSON using generate_descriptions for downstream dashboards.

Validate with experiments

Use the recommendations in Experiments & Benchmarks to design reproducible evaluations before moving to production.