parameters.md

Run Parameters

STREAMLINE parameters can be supplied through .cfg files, notebooks, or phase CLI flags. The recommended full-pipeline path is the config runner:

python run.py -c run_configs/local/uci_binary_hcc.cfg --dry_run
python run.py -c run_configs/local/uci_binary_hcc.cfg

The .cfg parameter names intentionally match the command-line names wherever possible. This page has two layers: a short set of essential parameters that most users need to set correctly, followed by a more exhaustive phase-by-phase reference. Essential parameters are repeated in the exhaustive reference so the reader can either skim from the top or look up a single phase later.

Use these terms consistently when reading the tables:

• Task type means the supervised learning problem: Binary, Multiclass, or Continuous. This is controlled by outcome_type.
• Execution mode means where jobs run: Serial, Parallel, Local, BashSLURM, BashLSF, or a named Dask cluster. This is controlled by run_cluster.
• Run path means how STREAMLINE is launched: a notebook, the full config runner, or an individual phase CLI.
• Report mode means whether P11 creates a standard training/CV report or a replication report.

Later phases can load values saved by earlier phases in metadata.pickle and run_commands.pickle. When a default below says it comes from metadata, the owning phase is named where possible. For example, outcome labels, feature types, CV counts, and one-hot settings are saved by P1; imputation, scaling, and SMOTE settings are saved by P2; feature-learning settings are saved by P3; feature-importance settings are saved by P4; modeling settings are saved by P6. Explicit .cfg or CLI values override remembered values.

Essential Parameters To Run STREAMLINE

These are the parameters most users should understand before starting a run. Start from one of the included files in run_configs/local/ or run_configs/hpc/, change these values, and use --dry_run to inspect the resolved phase calls before launching a full analysis.

Every Config Run

These parameters define the run itself. They belong in [run] for config files, or must be repeated on each individual phase CLI when running phases manually.

Parameter	Default value	Where to set it	Description
output_path	Required	[run]; every phase CLI	Parent folder where STREAMLINE writes the experiment output.
experiment_name	Required	[run]; every phase CLI	Name of the experiment folder created under output_path.
outcome_label	Class	[run], [p1]; repeated by later phase CLIs when needed	Outcome column in the input data. P1 records it for later phases.
outcome_type	P1 can infer; later phases should be explicit	[run], [p1], [p6], [p8], [p9], [p11]	Task type: Binary, Multiclass, or Continuous. Set this explicitly for paper or benchmark runs.
instance_label	None	[run], [p1]; repeated by later phase CLIs when needed	Optional row identifier column. P1 records it and excludes it from modeling.
n_splits	10	[run]; CV-aware phase CLIs	Number of cross-validation folds. Demo configs use 3 for speed.
run_cluster	Serial	[run]; every phase CLI	Execution mode. Use Parallel for local joblib multiprocessing, Local for local Dask, or BashSLURM/BashLSF for scheduler submission.
random_state	Phase-specific, often None or 0	[run]; stochastic phase CLIs	Seed for reproducible CV partitioning, imputation/SMOTE, feature learning, modeling, and ensembles.
phase_order	p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11	[phases]	Ordered list of phases for the config runner.
do_p1 through do_p11	True unless disabled	[phases]	Phase toggles. P10 also requires replication paths; P7 is skipped for continuous outcomes.

Phase-Specific Essentials

These parameters decide what data are analyzed, how features are handled, which models run, and which reports are produced. They are grouped by the phase that first uses or remembers them.

Parameter	Default value	Where to set it	Description
data_path	Required for P1	[p1]	Folder containing one or more input .csv, .tsv, or .txt datasets.
categorical_features	None	[p1]	Optional file listing categorical feature names. Recommended when feature types matter.
quantitative_features	None	[p1]	Optional file listing quantitative feature names. Recommended with categorical_features.
ignore_features	None	[p1]	Optional file or list of feature names to exclude before modeling.
partition_method	Stratified	[p1]	CV strategy. Continuous outcomes are forced to Random.
one_hot_encoding	True	[p1]	Expand non-binary categorical features in P1. If False, P6 only allows native-categorical models unless that guard is disabled.
scale_data	P2 remembered value, fallback True	[p2], [p8]	Applies scaling in P2 and records whether scaled data was used in summary/reporting.
impute_data	P2 remembered value, fallback True	[p2]	Enables missing-value imputation for CV train/test folds.
smote	P2 remembered value, fallback False	[p2]	Enables classification-only training-fold oversampling after imputation and scaling.
models	All available non-excluded P6 models	[p6]	Model IDs to train, such as NB,LR,DT,RF,CGB,HEROS,ExSTraCS. Demo configs use small model lists for speed.
model_params_json	None	[p6]	Optional JSON or Python-literal dictionary of model-specific overrides. See Model Parameter JSON.
scoring_metric	balanced_accuracy	[p6], [p8]	Primary modeling/evaluation metric. Use explained_variance for regression unless intentionally changing the regression metric.
metric_direction	maximize	[p6]	Optuna optimization direction. Use minimize only for loss/error metrics where lower is better.
n_trials	200	[p6]	Maximum Optuna trials per model/CV job.
timeout	900	[p6]	Maximum Optuna time budget in seconds per model/CV job.
training_subsample	0	[p6]	Optional cap on training rows for models that explicitly allow subsampling. 0 disables it.
skip_completed_models	False	[p6]	When True, P6 runs only missing or failed model/CV jobs. Default behavior reruns requested model jobs and overwrites artifacts.
rep_data_path	Required for P10	[p10]	Folder containing replication/external-validation datasets.
dataset_for_rep	Required for P10	[p10]	Original training dataset path used to identify the trained dataset output folder.
report_modes	standard for P11; demo configs use standard,replication	[p11]	Report types generated by the config runner.

CLI-only controls are prefixed with -- in the exhaustive reference. They are not written into .cfg files. Use them to choose a config file, dry-run a config, run only part of the phase order, list registry methods, or control saved run-command reuse.

Config Template Folders

The run_configs/ directory is organized by execution environment:

Folder	Use case	Included examples
run_configs/local/	Local serial, local joblib Parallel, and local Dask Local runs.	Binary HCC, multiclass student dropout, and regression Auto MPG demo configs.
run_configs/hpc/	Scheduler-oriented templates that use BashSLURM, BashLSF, or site-specific cluster settings.	cedars_slurm_hcc.cfg for Cedars/SLURM and upenn_lsf_hcc.cfg for UPenn/LSF starting points.

The original top-level demo configs are still kept for backward compatibility, but new examples should point users to the environment-specific subfolders.

For cluster-specific workflow details, including tmux, SLURM/LSF monitoring, and recovery after interrupted runs, see HPC and Cluster Runs.

Full Parameter Reference By Phase

Defaults below are the current runner defaults when a parameter is omitted. Where noted, later phases may load the value from experiment metadata saved by earlier phases.

Shared Run Parameters

Parameter	Default value	Description
output_path	Required	Parent folder for experiment outputs.
experiment_name	Required	Experiment folder name under output_path.
outcome_label	Class	Outcome column. Passed to P1 and reused by modeling, evaluation, comparison, replication, and reporting.
outcome_type	None in P1; later phases use metadata when possible	Learning task type: Binary, Multiclass, or Continuous.
instance_label	None	Optional row identifier column.
n_splits	10	Number of CV folds for CV-aware phases.
run_cluster	Serial	Execution mode. Parallel uses local joblib multiprocessing; Local uses a local Dask cluster; BashSLURM and BashLSF submit scheduler scripts.
queue	defq	Scheduler queue/partition for BashSLURM, BashLSF, or named cluster execution.
reserved_memory	4	Memory request in GB for submitted cluster jobs.
random_state	None in P1 and P6; metadata or 0 in several later phases	Seed for stochastic steps. Set explicitly for reproducible paper runs.
wait_for_cluster_completion	True for BashSLURM/BashLSF config runs	Makes the config runner wait for submitted cluster jobs to write completion markers before starting the next phase.
cluster_phase_timeout	86400	Maximum seconds to wait for a submitted cluster phase.
cluster_phase_poll_interval	30	Seconds between completion-marker checks for submitted cluster phases.

Config Runner Controls

These are command-line controls for python run.py -c ..., not .cfg keys.

Parameter	Default value	Description
--config, -c	Required	Path to a STREAMLINE .cfg or .ini file.
--dry_run	False	Print resolved phase runner calls without running phases.
--start_at	None	Start at a phase alias such as p4 or p6_modeling.
--stop_after	None	Stop after a phase alias such as p8 or p11.
--only	None	Run only a comma-separated set of phase aliases.
--skip	None	Skip a comma-separated set of phase aliases.
--log_level	INFO	Python logging level for the config runner.

Phase Toggles

Parameter	Default value	Description
phase_order	p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11	Phase order used by the config runner.
do_p1	True	Run P1 data exploration and processing.
do_p2	True	Run P2 imputation, scaling, and optional SMOTE.
do_p3	True	Run P3 feature learning.
do_p4	True	Run P4 feature importance.
do_p5	True	Run P5 feature selection.
do_p6	True	Run P6 modeling.
do_p7	True	Run P7 ensembles. P7 is skipped automatically for continuous outcomes.
do_p8	True	Run P8 summary statistics and plots.
do_p9	True	Run P9 dataset comparison. P9 skips itself when fewer than two datasets are available.
do_p10	True when replication paths are configured	Run P10 replication or external validation.
do_p11	True	Run P11 reporting.
enabled	True	Per-phase override for disabling an individual phase section, commonly used as enabled = False in [p7] for regression configs.
do_all	Not set	Old-style broad toggle that enables or disables all phases when present.
do_till_report	Not set	Old-style broad toggle for running phases through the standard report path.

P1 Data Process

Parameter	Default value	Description
data_path	Required	Folder containing raw input datasets, or omitted only when importing prebuilt CV datasets.
exclude_eda_output	None	Optional list of EDA outputs to skip, such as describe_csv or correlation.
match_label	None	Optional column label used when matching or harmonizing datasets.
ignore_features	None	Optional file or list of feature names to exclude.
categorical_features	None	Optional feature-name file for categorical variables.
quantitative_features	None	Optional feature-name file for quantitative variables.
top_features	20	Number of top features shown in applicable P1 summaries.
categorical_cutoff	10	If feature-type files are absent, features with at most this many unique values may be treated as categorical.
sig_cutoff	0.05	Statistical significance threshold used in P1 analyses.
featureeng_missingness	0.5	Missingness threshold for creating missingness indicator features.
cleaning_missingness	0.5	Missingness threshold for removing high-missingness features or instances.
correlation_removal_threshold	1.0	Correlation threshold for removing highly correlated features. 1.0 effectively disables correlation removal.
partition_method	Stratified	CV partitioning strategy. Continuous outcomes are forced to Random.
show_plots	False	Display P1 plots interactively. Usually False for batch runs.
one_hot_encoding	True	Expand non-binary categorical features during P1 processing.
cv_provided	False	Import existing CV train/test files instead of creating CV splits from raw datasets.
cv_input_root	None	Root folder containing prebuilt <dataset>/CVDatasets folders when cv_provided=True.
enable_plots	False	Master toggle for optional P1 plot generation.
plot_missingness	False	Generate missingness plots.
plot_class_counts	False	Generate outcome/class count plots.
plot_correlation	False	Generate correlation plots.
correlation_plot_max_features	200	Maximum number of features included in correlation plots.
plot_univariate	False	Generate univariate feature analysis plots.
univariate_top_k	20	Number of top univariate features to display.
plot_anomalies	False	Generate anomaly/outlier plots when available.
force	False	Overwrite existing P1 outputs. Demo configs set this to True for easy reruns.

P2 Impute, Scale, And Balance

Parameter	Default value	Description
scale_data	P2 saved metadata, fallback True	Scale features using the selected scaler.
impute_data	P2 saved metadata, fallback True	Impute missing feature values.
multi_impute	P2 saved metadata, fallback False	Use multivariate imputation for quantitative features when supported.
overwrite_cv	True	Rewrite CV train/test files with P2 outputs.
outcome_label	P1 saved metadata, fallback Class	Outcome column.
outcome_type	P1 saved metadata, fallback None	Learning task type.
instance_label	P1 saved metadata, fallback None	Optional row identifier column.
random_state	P1/P2 saved metadata, fallback 0	Seed for stochastic imputers or SMOTE.
imputer_id	P2 saved metadata, fallback None	Registry imputer ID. None uses the phase default.
imputer_params	P2 saved metadata, fallback {}	Dictionary of imputer parameters.
scaler_id	P2 saved metadata, fallback None	Registry scaler ID. None uses the phase default.
scaler_params	P2 saved metadata, fallback {}	Dictionary of scaler parameters.
smote	P2 saved metadata, fallback False	Apply classification-only oversampling to training folds after imputation and scaling.
smote_method	P2 saved metadata, fallback auto	auto, smote, or smotenc. auto uses SMOTENC when categorical features are present.
smote_sampling_strategy	P2 saved metadata, fallback auto	Sampling strategy passed to imbalanced-learn.
smote_k_neighbors	P2 saved metadata, fallback 5	Neighbor count passed to SMOTE or SMOTENC.
--list-imputers	False	CLI-only utility: list discovered imputer registry IDs and exit.
--list-scalers	False	CLI-only utility: list discovered scaler registry IDs and exit.

P3 Feature Learning

Parameter	Default value	Description
learner_id	P3 saved metadata, fallback pca	Feature learner registry ID.
learner_params	P3 saved metadata, fallback {}	Dictionary of learner parameters.
feature_namespace	P3 saved metadata, fallback FL_PCA	Prefix/namespace for learned feature names.
keep_original_features	P3 saved metadata, fallback True	Keep original features alongside learned features.
overwrite_cv	True	Rewrite CV train/test files with P3 outputs.
outcome_label	P1 saved metadata, fallback Class	Outcome column.
instance_label	P1 saved metadata, fallback None	Optional row identifier column.
random_state	P1/P3 saved metadata, fallback 0	Seed for stochastic learners.
--list-learners	False	CLI-only utility: list discovered feature-learning registry IDs and exit.

P4 Feature Importance

Parameter	Default value	Description
models	P4 saved metadata, fallback all registered FI methods	Feature-importance methods to run, such as mutualinformation,multiswrfdb.
models_params	P4 saved metadata, fallback ReBATE n_jobs=1 defaults where applicable	Per-method parameter dictionary. STREAMLINE injects saved categorical feature indexes for ReBATE methods.
top_k	P4 saved metadata, fallback None	Optional top-k selector control for model-specific selected outputs.
threshold	P4 saved metadata, fallback None	Optional score threshold for model-specific selected outputs.
keep_original_features	P4 saved metadata, fallback False	Keep original features in selected-output artifacts when generated.
overwrite_cv	True	Overwrite P4 model-specific outputs. Shared CV files are not mutated by P4.
outcome_label	P1 saved metadata, fallback Class	Outcome column.
outcome_type	P1 saved metadata, fallback None	Learning task type passed to compatible FI methods.
instance_label	P1 saved metadata, fallback None	Optional row identifier column.
random_state	P1/P4 saved metadata, fallback 0	Seed for stochastic FI methods.
instance_subset	P4 saved metadata, fallback None	Optional row cap for expensive FI methods. No subsampling is used when None.
--list-models	False	CLI-only utility: list discovered feature-importance methods and exit.

P5 Feature Selection

Parameter	Default value	Description
algorithms	auto	FI algorithms considered by the selector. auto discovers completed P4 outputs.
n_splits	10	Number of CV folds expected in FI outputs. Usually inherited from [run].
outcome_label	P1 saved metadata, fallback Class	Outcome column.
instance_label	P1 saved metadata, fallback None	Optional row identifier column.
max_features_to_keep	2000	Upper bound on selected features after combining FI rankings.
filter_poor_features	True	Remove features with consistently poor or zero FI evidence.
overwrite_cv	False	Overwrite P5 selected CV outputs.
selector_id	default	Feature selector registry ID.
selector_params	{}	Dictionary of selector parameters.
export_scores	True	Write feature-selection score summaries.
top_features	20	Number of top features shown in P5 plots/summaries.
show_plots	False	Display P5 plots interactively.
strict_discovery	False	Require all expected CV FI files for an algorithm during auto discovery.
--list-algorithms	False	CLI-only utility: list available/discovered FI algorithms and exit.

P6 Modeling

Parameter	Default value	Description
outcome_type	None; resolved from model_type, otherwise Binary	Modeling task: Binary, Multiclass, or Continuous. The config runner fills this from [run] when available.
model_type	None	Backward-compatible alias for outcome_type; prefer outcome_type in new configs.
models	All available non-excluded models for the task	Model registry IDs. eLCS is excluded from default discovery.
model_params_json	None	Optional JSON or Python-literal mapping of model IDs to parameter overrides. See Model Parameter JSON.
calibrate	False	Enable probability calibration for classification models.
calibrate_method	sigmoid	Calibration method, usually sigmoid or isotonic.
calibrate_cv	5	Internal CV folds used for calibration.
scoring_metric	balanced_accuracy	Optuna/evaluation metric. Regression configs should use a regression metric such as explained_variance.
metric_direction	maximize	Optuna optimization direction.
n_trials	200	Maximum Optuna trials per model/CV job.
timeout	900	Maximum Optuna seconds per model/CV job.
training_subsample	0	Optional training subset size for models with subsampling_allowed=True; 0 disables subsampling.
uniform_fi	False	Use uniform permutation FI handling when supported.
save_plot	False	Save model-level plots generated during modeling.
skip_completed_models	False	When True, run only failed or missing model/CV jobs. When False, rerun requested jobs and overwrite artifacts.
bypass_one_hot_for_native_models	True	Allow the native categorical model path when P1 was run with one_hot_encoding=False.
native_categorical_models	CGB,ExSTraCS	Allowed native-categorical model IDs when one-hot encoding is bypassed.
--list_models	False	CLI-only utility: list default model IDs for the selected task and exit.
--list_models_all	False	CLI-only utility: list all registered model IDs for all tasks and exit.

P6 records Optuna trial accounting in model outputs so reports can show how many trials actually ran within the requested budget.

By default, P6 reruns the requested model/CV jobs and overwrites existing model artifacts. Use skip_completed_models = True in a config file, or --skip_completed_models 1 on the P6 CLI, when you want recovery behavior that skips completed job_model_* markers and runs only failed or missing jobs.

P7 Ensembles

P7 is classification-only. The config runner skips P7 automatically for continuous outcomes.

Parameter	Default value	Description
ensembles	hard_voting,soft_voting,stack_lr	Ensemble registry IDs.
base_models	None	Base model predictions to combine. None lets P7 discover compatible model outputs.
meta_train_source	train	Source for stacking meta-training data: train or test.
calibrate	False	Enable calibration for ensemble probabilities when supported.
calibrate_method	sigmoid	Calibration method, usually sigmoid or isotonic.
calibrate_cv	5	Internal CV folds used for calibration.
random_state	0	Seed for stochastic ensemble behavior.
--list_ensembles	False	CLI-only utility: list ensemble registry IDs and exit.

P8 Summary Statistics

Parameter	Default value	Description
outcome_type	P1/P6 saved metadata, fallback Binary	Learning task used to choose classification or regression summaries.
scoring_metric	balanced_accuracy	Primary metric label used in summaries.
metric_weight	balanced_accuracy	Metric used to weight composite model FI plots. Continuous outcomes default to explained_variance if an incompatible metric is supplied.
top_features	40	Number of top features shown in composite FI visualizations.
sig_cutoff	0.05	Statistical significance threshold for comparisons.
scale_data	True	Metadata/reporting flag indicating whether scaled data are being summarized.
exclude_plots	None or empty string	Comma-separated plots to skip, such as plot_ROC,plot_PRC,plot_FI_box,plot_metric_boxplots.
show_plots	False	Display P8 plots interactively.
include_ensembles	True	Include P7 ensemble outputs in summaries when present.
multiclass_average	micro	Multiclass averaging mode for ROC/PRC summaries: micro or macro.

P9 Compare Datasets

Parameter	Default value	Description
outcome_label	Class	Outcome column.
outcome_type	Binary	Learning task type.
instance_label	None	Optional row identifier column.
sig_cutoff	0.05	Statistical significance threshold for between-dataset comparisons.
show_plots	False	Display P9 plots interactively.

P9 compares datasets within the same experiment and writes a skipped marker when fewer than two dataset folders with CVDatasets/ are present.

P10 Replication

Parameter	Default value	Description
rep_data_path	Required	Folder containing external replication datasets.
dataset_for_rep	Required	Original training dataset path used to identify the trained dataset output folder.
outcome_label	P1 saved metadata	Optional override for the outcome column.
instance_label	P1 saved metadata	Optional override for the row identifier column.
match_label	None	Optional label used to match or harmonize replication inputs.
exclude_plots	None	Comma-separated plots to skip, such as plot_ROC, plot_PRC, plot_metric_boxplots, plot_FI_box, or feature_correlations.
show_plots	False	Display replication plots interactively.

P11 Reporting

Parameter	Default value	Description
experiment_path	Required unless output_path and experiment_name are provided	Direct path to the experiment output folder.
output_path	Required unless experiment_path is provided	Parent output folder.
experiment_name	Required unless experiment_path is provided	Experiment folder name.
reporting_dir	None	Optional directory for report artifacts. None uses the standard experiment reporting folders.
report_modes	standard in config runner unless set	Config-runner convenience parameter for generating multiple report modes, such as standard,replication.
report_mode	standard	Single report mode: standard or replication.
outcome_label	Class in runner, often loaded from metadata/report data	Outcome column used in report labels.
outcome_type	Binary in runner, often loaded from metadata/report data	Learning task type used in report labels and metric filtering.
instance_label	None	Optional row identifier column.
make_pdf	True	Export a PDF report.
enable_plots	True	Generate missing report plots when possible.
reuse_existing_figures	True	Reuse existing report figure PNGs when available. Set to False to regenerate report figures.

Saved Run Command Controls

All phase CLIs support these run-command controls.

Flag	Default value	Description
--ignore_saved_run_command	False	Ignore run_commands.pickle for this run.
--no_update_saved_run_command	False	Do not update run_commands.pickle after the run.

Use these flags when you want to run a phase with explicit command-line values instead of reusing arguments saved from a previous run.