auto-ml-pipeline

name	auto-ml-pipeline
description	Automated machine learning pipeline for tabular data. Full lifecycle: preprocessing, feature engineering, model selection, hyperparameter tuning, evaluation, and export.
version	1.0.0

AutoML Pipeline Skill

Overview

Automated machine learning pipeline for tabular data. Handles the full lifecycle: preprocessing → feature engineering → model selection → hyperparameter tuning → evaluation → export.

Architecture

AutoMLClassifier / AutoMLRegressor
    │
    ├── Preprocessor        (missing values, outliers, encoding, scaling)
    ├── FeatureEngineer     (polynomial features, variance filter, mutual info selection)
    ├── ModelSelector       (train multiple models, rank by CV score)
    ├── HyperparameterTuner (Optuna Bayesian optimization)
    ├── Evaluator           (classification/regression metrics)
    ├── Explainer           (SHAP-based feature importance)
    └── ModelExporter       (joblib/pickle/ONNX serialization)

Core Classes

Class	File	Purpose
AutoMLClassifier	automl/pipeline.py	Classification pipeline entry point
AutoMLRegressor	automl/pipeline.py	Regression pipeline entry point
Preprocessor	automl/preprocessor.py	Impute, remove outliers, encode, scale
FeatureEngineer	automl/feature_engineer.py	Polynomial expansion, variance threshold, feature selection
ModelSelector	automl/model_selector.py	Train baseline models (linear, tree, SVM, ensemble, MLP)
HyperparameterTuner	automl/tuner.py	Optuna-based search with per-model search spaces
Evaluator	automl/evaluator.py	Metrics: accuracy, F1, ROC-AUC, R², RMSE, etc.
Explainer	automl/explainer.py	SHAP TreeExplainer / KernelExplainer / LinearExplainer
ModelExporter	automl/exporter.py	Save/load with metadata JSON

Pipeline Flow (fit)

1. Preprocess — impute missing → remove IQR outliers → encode categoricals → scale numerics
2. Feature Engineer — polynomial expansion → variance threshold → mutual info selection
3. Model Selection — cross-validate all enabled model categories, pick best by CV score
4. Tuning — Optuna optimize on the best model class (half of time_limit)
5. Evaluate — compute metrics on held-out or full data

Usage Pattern

from automl import AutoMLClassifier, AutoMLRegressor

# Classification
clf = AutoMLClassifier(time_limit=300, n_trials=50, cv_folds=5)
clf.fit(X_train, y_train, X_val=X_val, y_val=y_val)
preds = clf.predict(X_test)
metrics = clf.evaluate(X_test, y_test)
importance = clf.explain(X_test)
clf.export_model("./models", "my_classifier")

# Regression — same API with AutoMLRegressor

Configuration

Load via config_path="configs/default_config.yaml" or override flags in code:

automl.model_selector_.include_linear = False
automl.model_selector_.include_tree = True   # RandomForest, DecisionTree, GradientBoosting only
automl.model_selector_.include_svm = False
automl.model_selector_.include_ensemble = False
automl.model_selector_.include_neural_network = False

Config sections: preprocessing, feature_engineering, models, optimization, evaluation.

Model Categories (ModelSelector)

Flag	Classification Models	Regression Models
include_linear	LogisticRegression	LinearRegression, Ridge, Lasso, ElasticNet
include_tree	RandomForest, DecisionTree, GradientBoosting	Same + Reg variants
include_svm	SVC (probability=True)	SVR
include_ensemble	VotingClassifier (RF+GB)	VotingRegressor
include_neural_network	MLPClassifier	MLPRegressor

Tuner Search Spaces

Each model class has a predefined Optuna search space in HyperparameterTuner._define_search_space():

• RandomForest — n_estimators, max_depth, min_samples_split/leaf, max_features, bootstrap
• GradientBoosting — n_estimators, learning_rate (log), max_depth, subsample, max_features
• DecisionTree — max_depth, min_samples_split/leaf, max_features
• Linear models — C (log), penalty (l1/l2), solver
• SVM — C (log), kernel, gamma
• MLP — hidden_layer_sizes (dynamic depth), alpha (log), learning_rate_init (log), batch_size

Key Constraints

• Input: pd.DataFrame or np.ndarray for X, np.ndarray/list for y
• Output: np.ndarray predictions
• No hardcoded paths — all paths are relative/configurable
• SHAP is optional (pip install shap)
• ONNX export requires onnx + skl2onnx (optional)
• Uses numpy types internally — _make_serializable() converts to native Python for JSON

Tests

Located in tests/: test_preprocessor.py, test_feature_engineer.py, test_model_selector.py, test_tuner.py, test_pipeline.py.

Dependencies

Core: scikit-learn, pandas, numpy, optuna, pyyaml, joblib Optional: shap, matplotlib, onnx, skl2onnx