Changelog

This changelog summarizes notable STREAMLINE changes in newest-first order. Older public release notes are based on the GitHub Releases entries, with minor wording cleanup for readability.

v1.0.0 - Main Release

STREAMLINE v1.0.0 is a major reorganization and expansion of STREAMLINE into a broader supervised tabular AutoML pipeline. It keeps the original transparent, end-to-end design while adding first-class support for more task types, clearer phase boundaries, registry-based extension points, updated notebooks, Sphinx documentation, and broader testing.

Added

• Added explicit P1-P11 phase modules for data processing, imputation/scaling and balancing, feature learning, feature importance, feature selection, modeling, classification ensembles, summary statistics, dataset comparison, replication, and reporting.

• Added run-command pickle support so repeated phase calls can reuse the parameters actually used in previous phases, while still allowing user overrides.

• Added binary classification, multiclass classification, and regression as supported task types across core pipeline phases.

• Added optional probability calibration for classification base models and ensembles, with configurable calibration method and CV folds.

• Added UCI demo datasets and held-out replication splits for binary, multiclass, and regression examples.

• Added optional SMOTE/SMOTENC balancing after Phase 2 imputation and scaling.

• Added P3 feature learning as a dedicated phase, including PCA-style learned feature outputs and manifests.

• Added MultiSWRFDB and MultiSWRFDB* feature-importance methods.

• Added expanded task-specific model registries for binary classification, multiclass classification, and regression workflows.

• Added multiclass Decision Tree support.

• Added HEROS and optional TabPFN wrappers, including token-aware TabPFN skip behavior when TABPFN_TOKEN is unavailable.

• Added Optuna trial accounting so model outputs and reports can distinguish requested hyperparameter budgets from completed trials.

• Added local Parallel execution in addition to Serial, local Dask through Local, and supported cluster submission modes.

• Added standard and replication report-mode separation, experiment-name-aware report titles and filenames, resolved/default parameter display, and structured report_data.json output for report debugging.

• Added feature-learning and feature-selection summary tables modeled after the data-processing and feature-engineering summary table style.

• Added Sphinx/autodoc documentation and GitHub test workflows for Python 3.10, 3.11, and 3.12.

Changed

• Reorganized code around registries for Phase 2 preprocessing, Phase 3 feature learning, Phase 4 feature importance, Phase 5 selection, Phase 6 models, and Phase 7 ensembles.

• Updated the existing configuration-file workflow around .cfg files with aligned parameter names, UCI demo configs, dry-run support, phase selection flags, and matching notebook parameters.

• Standardized parameter names across config files, command-line arguments, notebooks, and saved run-command metadata.

• Updated Phase 1 task handling so user-specified outcome_type is respected instead of being inferred only from the number of unique outcome values.

• Updated Phase 4 so feature-importance methods write scores without mutating shared CV datasets.

• Updated scikit-rebate integration to pass STREAMLINE categorical feature indexes to ReBATE methods.

• Updated Phase 5 feature selection so rankings can be combined across all feature-importance methods that were run.

• Reworked Phase 6 modeling around clearer dataset/model/CV execution units for serial, parallel, Dask, and cluster runs.

• Updated ExSTraCS categorical initialization using supported discrete_attribute_limit and specified_attributes parameters.

• Updated native categorical handling for compatible algorithms such as CatBoost/CGB and ExSTraCS.

• Updated reports and summary statistics so metric names, no-skill baselines, curves, and plots are task-aware rather than binary-only.

• Updated replication and reporting flows to handle binary, multiclass, and regression outputs with task-specific metrics and plots.

• Updated Google Colab and local Jupyter notebooks for binary, multiclass, regression, and custom dataset workflows.

Fixed

• Fixed replication behavior that could otherwise zero-fill learned/PCA feature columns instead of respecting saved feature-learning artifacts.

• Fixed Phase 4 shared CV file mutation and parallel race risks.

• Fixed training_subsample timing so subsampling affects the intended model training workflow.

• Fixed multiclass XGB/LGB objective handling and binary-only assumptions.

• Fixed binary ensemble confusion-metric extraction so confusion matrices are not dropped before TP/TN/FP/FN-derived metrics are calculated.

• Fixed P1 bash/job submission path issues.

• Fixed report wording where unspecified settings should instead show the actual default used or indicate that a phase was not run.

• Fixed no-skill ROC/PR legend notes and label-aware baseline handling for multiclass or non-stratified/random CV settings.

v0.3.4-beta - 2023-09-28

Changed

• Improved PDF report formatting so first-page content is clearer and reports handle larger numbers of analyzed datasets more gracefully.

Fixed

• Fixed an edge-case failure when running multiple separate replication and replication-report phases in legacy mode.

v0.3.3-beta - 2023-09-23

Added

• Added invariant feature removal during the C2 cleaning step of data processing.

• Added matching replication behavior so features removed during Phase 1 cleaning are also removed when replication data are processed.

Changed

• Updated replication PDF report content to simplify the data-processing report.

• Updated first-page PDF report text sizing.

Fixed

• Fixed algorithm ordering in notebook and Colab figures.

• Fixed unseen binary categorical values during replication by converting them to missing values instead of adding incompatible new features.

• Fixed engineered missingness feature naming.

• Fixed legacy cluster mode when categorical or quantitative feature files were not specified.

v0.3.2-beta - 2023-09-13

Changed

• Updated legacy run mode so submitted jobs can be launched and the script can exit instead of waiting for all jobs to complete.

• Updated the STREAMLINE schematic.

• Updated PDF summary file naming.

• Added documentation describing how to check job status.

Fixed

• Fixed command-line argument passing for legacy run mode.

v0.3.1-beta - 2023-09-07

Changed

• Ordered plot legends, including composite feature-importance plot legends, alphabetically by full model name.

Fixed

• Fixed replication imputation when a feature had no missing values during training but did have missing values in the replication dataset. The replication phase now applies a simple fallback imputation strategy: mean for quantitative features and mode for categorical features.

v0.3.0-beta - 2023-08-06

Added

• Added Dask jobqueue support for running STREAMLINE across several HPC cluster systems.

• Expanded Phase 1 from exploratory analysis into numerical encoding, automated cleaning, feature engineering, and a second processed-data EDA pass.

• Added numerical encoding maps for binary text-valued features.

• Added categorical and quantitative feature path parameters, plus output files documenting final feature-type handling.

• Added missingness feature engineering with output documentation.

• Added feature and instance cleaning based on missingness.

• Added one-hot encoding for categorical features with three or more values.

• Added highly correlated feature removal with output documentation.

• Added DataProcessSummary.csv to track feature, feature-type, instance, class, and missing-value counts through data-processing steps.

• Added replication processing parity so replication data are transformed to match the target dataset feature space.

• Added command-line support for running the whole pipeline as a single command.

• Added configuration-file support for command-line runs.

• Added class-based modeling algorithm modules to make adding compatible scikit-learn-style classifiers easier.

• Added Elastic Net as an included modeling algorithm.

• Added expanded Colab workflows with repository download, easy/manual run modes, user data selection, and output download/report display support.

• Added custom HCC demo and replication datasets designed to exercise automatic data cleaning, feature engineering, and replication behavior.

Changed

• Reorganized repository hierarchy, file names, output names, and phase groupings.

• Updated the STREAMLINE schematic to match the reorganized phase structure.

• Reverted model feature-importance plot sorting/presentation from median back to mean to avoid confusing demo behavior with small CV counts.

• Updated feature correlation heatmap colors, non-redundant triangle display, feature-name scaling, and large-feature-count behavior.

• Added FeatureCorrelations.csv output.

• Reformatted PDF summaries to reorganize first-page run parameters, include version text, and add data-processing/count summaries.

• Added test run and score outputs to univariate analysis files.

• Updated Jupyter and useful notebooks for the reorganized framework.

v0.2.5-beta - 2022-06-24

Fixed

• Added a catch to prevent statistical-comparison result failures in specific edge cases.

• Cleaned up old commented code.

v0.2.4-beta - 2022-06-15

Changed

• Switched feature-importance figure summaries from mean to median at a collaborator’s recommendation.

• Added median algorithm performance summaries and median performance values to PDF summaries.

• Updated statistical significance output to present medians, matching the non-parametric statistical comparisons more closely.

Fixed

• Fixed a no-missing-data edge case where imputation was enabled but no imputation file existed.

• Updated model application so replication data can be loaded from .csv and .txt files.

v0.2.3-beta - 2022-05-20

Added

• Confirmed stable serial command-line functionality on Linux.

Changed

• Marked the beta line as stable and fully functional based on testing and user feedback after the alpha releases.

v0.2.2-beta - 2022-05-19

Changed

• Added composite feature-importance plot weighting by balanced accuracy and ROC AUC.

• Removed the None option for maximum features in feature selection.

• Updated Logistic Regression Optuna search behavior to avoid invalid hyperparameter combinations.

• Enforced Optuna 2.0.0 for hyperparameter-optimization figure generation and added error handling so plotting issues do not fail an entire STREAMLINE run.

• Updated notebooks for the beta fixes.

Fixed

• Fixed composite feature importance when only one algorithm was used.

• Fixed major issues preventing certain phases from running serially from the command line.

• Fixed first-page PDF summary formatting.

v0.2.1-beta - 2022-05-17

Changed

• Moved the codebase into the streamline package folder and updated imports accordingly.

• Updated default Optuna run parameters.

• Documented that complete reproducibility is not guaranteed when Optuna is run in parallel.

• Rounded scaled CV data to seven decimal places to reduce floating-point reproducibility drift after scaling.

v0.2.0-beta - 2022-05-14

Added

• Published the first beta release after initial alpha testing across multiple platforms and Anaconda installations.

Changed

• Marked STREAMLINE ready for external use while noting that untested configurations could still expose issues.

• Added guidance for users to report run mode, Anaconda version, and errors when issues arise.

• Added an early request for users applying STREAMLINE in publications to check the repository for the current citation reference.

v0.1.3-alpha - 2022-05-12

Changed

• Updated README installation instructions.

• Updated the default setting for model feature-importance estimation.

• Recorded testing against the then-current Linux Anaconda version.

v0.1.2-alpha - 2022-05-12

Fixed

• Fixed an Anaconda/scipy compatibility issue in exploratory analysis.

v0.1.1-alpha - 2022-05-12

Fixed

• Replaced deprecated scipy.interp() usage with numpy.interp().

v0.1.0-alpha - 2022-05-12

Added

• Published the first stable, bug-tested STREAMLINE implementation, inheriting much of its underlying code from AutoMLPipe-BC.

Notes

• This alpha was tested only with the specified Anaconda and package versions, and only on Windows and Linux.