Machine learning

scikit-learn datasets, models, evaluation and tuning.

15 nodes. Right-click any node in the editor to read this documentation in the app.

ML / Data

🔄 Apply Transformer

id sklearn_transform · ML / Data · Python export: yes

Fit a transformer on the TRAINING rows only, then transform both training and test rows (so nothing about the test data leaks into preprocessing). Outputs a new splits dict; y is unchanged.

Inputs

Port	Type	Description
transformer	Transformer	Unfitted transformer from a Transformer node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
splits	Splits	Splits with X_train / X_test transformed.

🌸 Sample Dataset

id sklearn_load_dataset · ML / Data · Python export: yes

One of scikit-learn's built-in datasets as a DataFrame: the feature columns plus a 'target' column. iris, wine, breast_cancer and digits are classification problems; diabetes is regression. Ships with scikit-learn, so it works offline.

Outputs

Port	Type	Description
dataframe	DataFrame	Feature columns plus a 'target' column.

Fields

Field	Type	Default	Choices
Dataset	select	iris	iris, wine, breast_cancer, diabetes, digits

✂️ Train/Test Split

id sklearn_train_test_split · ML / Data · Python export: yes

Split a DataFrame into training and test sets. Name the target column; every other column becomes a feature. A fixed random seed makes the split reproducible (use a negative number for a different split each run); 'Stratify' keeps class proportions equal in both sets.

Inputs

Port	Type	Description
dataframe	DataFrame	Table with the features and the target column.

Outputs

Port	Type	Description
splits	Splits	Dict of X_train, X_test, y_train, y_test (pandas objects).

Fields

Field	Type	Default	Choices
Target column	text	target
Test fraction	float	0.2
Random seed (negative = random)	int	42
Stratify by target	checkbox	false
Shuffle before splitting	checkbox	true

🔧 Transformer

id sklearn_transformer · ML / Data · Python export: yes

Configure a preprocessing step: standard / min-max / robust scaling, missing-value imputation (mean, median, most frequent), or PCA. Nothing is fitted here — wire it into Apply Transformer. 'Components' applies to PCA only; 'Extra parameters' is an optional JSON object of further scikit-learn arguments.

Outputs

Port	Type	Description
transformer	Transformer	Unfitted scikit-learn transformer.

Fields

Field	Type	Default	Choices
Transformer	select	standard_scaler	standard_scaler, minmax_scaler, robust_scaler, impute_mean, impute_median, impute_most_frequent, pca
Components (PCA only)	int	2
Extra parameters (JSON)	text

ML / Evaluation

🔲 Confusion Matrix

id sklearn_confusion_matrix · ML / Evaluation · Python export: yes

For a fitted classifier: how often each actual class was predicted as each class (rows = actual, columns = predicted). The diagonal is the correct predictions.

Inputs

Port	Type	Description
model	FittedModel	A fitted model — from Fit or a Hyper-parameter Search node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
matrix	DataFrame	Counts, rows = actual class, columns = predicted class.

Fields

Field	Type	Default	Choices
Rows	select	test	test, train

🔁 Cross-Validate

id sklearn_cross_val · ML / Evaluation · Python export: yes

k-fold cross-validation of an (unfitted) estimator on the TRAINING rows: a more reliable estimate than a single split. Outputs one score per fold; scoring 'default' uses the model's own (accuracy for classifiers, R² for regressors).

Inputs

Port	Type	Description
estimator	Estimator	Unfitted model from a Classifier, Regressor or Clusterer node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
scores	DataFrame	One row per fold: fold, score.

Fields

Field	Type	Default	Choices
Folds	int	5
Scoring	select	default	default, accuracy, f1_weighted, roc_auc, r2, neg_mean_squared_error, neg_mean_absolute_error

📊 Evaluate

id sklearn_evaluate · ML / Evaluation · Python export: yes

Score a fitted model. Classifiers: accuracy, precision, recall and F1 (class-weighted) plus ROC-AUC when the model gives probabilities. Regressors: R², MAE, MSE and RMSE. Outputs a metric / value table (plot it with Bar Chart).

Inputs

Port	Type	Description
model	FittedModel	A fitted model — from Fit or a Hyper-parameter Search node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
metrics	DataFrame	Two columns: metric, value.

Fields

Field	Type	Default	Choices
Rows	select	test	test, train

⭐ Feature Importance

id sklearn_feature_importance · ML / Evaluation · Python export: yes

Which features drive a fitted model? Uses the model's feature_importances_ (forests, boosting, trees) or the magnitude of its coefficients (linear models, averaged over classes). Sorted, most important first.

Inputs

Port	Type	Description
model	FittedModel	A fitted model — from Fit or a Hyper-parameter Search node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
importance	DataFrame	Columns: feature, importance.

ML / Models

🧠 Classifier

id sklearn_classifier · ML / Models · Python export: yes

Configure a classification model: logistic regression, random forest, gradient boosting, SVM, k-NN or decision tree. Nothing is trained until Fit. Only the parameters that apply to the chosen model are used: 'Trees' → random forest / gradient boosting; 'Max depth' → forests, boosting, decision tree (0 = no limit); 'C' → logistic regression / SVM; 'Neighbours' → k-NN; 'Learning rate' → gradient boosting; 'Kernel' → SVM. 'Extra parameters' is an optional JSON object of further scikit-learn arguments, e.g. {"min_samples_leaf": 3}.

Outputs

Port	Type	Description
estimator	Estimator	Unfitted scikit-learn model — wire into Fit, Cross-Validate or Search.

Fields

Field	Type	Default	Choices
Model	select	random_forest	logistic_regression, random_forest, gradient_boosting, svm, knn, decision_tree
Trees	int	100
Max depth (0 = no limit)	int	0
C (regularisation)	float	1.0
Neighbours	int	5
Learning rate	float	0.1
Kernel	select	rbf	rbf, linear, poly, sigmoid
Random seed (negative = random)	int	42
Extra parameters (JSON)	text

🫧 Clusterer

id sklearn_clusterer · ML / Models · Python export: yes

Configure an unsupervised clustering model: k-means, DBSCAN, agglomerative or Gaussian mixture. 'Clusters' applies to k-means, agglomerative and the mixture; 'Eps' / 'Min samples' to DBSCAN. Fit it on the features (labels are ignored), then use Predict.

Outputs

Port	Type	Description
estimator	Estimator	Unfitted scikit-learn model — wire into Fit, Cross-Validate or Search.

Fields

Field	Type	Default	Choices
Algorithm	select	kmeans	kmeans, dbscan, agglomerative, gaussian_mixture
Clusters	int	3
Eps (DBSCAN)	float	0.5
Min samples (DBSCAN)	int	5
Random seed (negative = random)	int	42
Extra parameters (JSON)	text

🏋️ Fit

id sklearn_fit · ML / Models · Python export: yes

Train a model on the training rows (X_train, and y_train when the model is supervised). Works on a copy, so the estimator node can feed other nodes unchanged. Outputs the fitted model.

Inputs

Port	Type	Description
estimator	Estimator	Unfitted model from a Classifier, Regressor or Clusterer node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
model	FittedModel	The trained model — wire into Predict, Evaluate, Feature Importance.

🔮 Predict

id sklearn_predict · ML / Models · Python export: yes

Run a fitted model on the test (or training) rows and output its predictions as a Series indexed like the rows. For clusterers these are the cluster labels; DBSCAN and agglomerative clustering can only label their training rows.

Inputs

Port	Type	Description
model	FittedModel	A fitted model — from Fit or a Hyper-parameter Search node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
predictions	Series	One prediction per row, named 'prediction'.

Fields

Field	Type	Default	Choices
Rows	select	test	test, train

📈 Regressor

id sklearn_regressor · ML / Models · Python export: yes

Configure a regression model: linear, ridge, lasso, random forest, gradient boosting, SVR, k-NN or decision tree. Nothing is trained until Fit. 'Alpha' applies to ridge / lasso. Only the parameters that apply to the chosen model are used: 'Trees' → random forest / gradient boosting; 'Max depth' → forests, boosting, decision tree (0 = no limit); 'C' → logistic regression / SVM; 'Neighbours' → k-NN; 'Learning rate' → gradient boosting; 'Kernel' → SVM. 'Extra parameters' is an optional JSON object of further scikit-learn arguments, e.g. {"min_samples_leaf": 3}.

Outputs

Port	Type	Description
estimator	Estimator	Unfitted scikit-learn model — wire into Fit, Cross-Validate or Search.

Fields

Field	Type	Default	Choices
Model	select	random_forest	linear_regression, ridge, lasso, random_forest, gradient_boosting, svr, knn, decision_tree
Alpha (ridge / lasso)	float	1.0
Trees	int	100
Max depth (0 = no limit)	int	0
Neighbours	int	5
Learning rate	float	0.1
C (SVR)	float	1.0
Kernel (SVR)	select	rbf	rbf, linear, poly, sigmoid
Random seed (negative = random)	int	42
Extra parameters (JSON)	text

ML / Tuning

🎛️ Hyper-parameter Search

id sklearn_search · ML / Tuning · Python export: yes

Find the best settings by cross-validation on the TRAINING rows. Give a JSON grid of parameter names to lists of values, e.g. {"n_estimators": [50, 100], "max_depth": [3, 5, 10]} — use the scikit-learn parameter names of the estimator. 'grid' tries every combination; 'random' tries a random sample of them. The output is the fitted search: it acts as the best model (wire it into Predict / Evaluate) and into Search Results for the full table.

Inputs

Port	Type	Description
estimator	Estimator	Unfitted model from a Classifier, Regressor or Clusterer node.
splits	Splits	Train/test data — from Train/Test Split or Apply Transformer.

Outputs

Port	Type	Description
search	FittedModel	Fitted search; predicts with the best parameters found.

Fields

Field	Type	Default	Choices
Strategy	select	grid	grid, random
Parameter grid (JSON)	textarea	{"n_estimators": [50, 100], "max_depth": [3, 5,…
CV folds	int	5
Scoring	select	default	default, accuracy, f1_weighted, roc_auc, r2, neg_mean_squared_error, neg_mean_absolute_error
Samples (random strategy)	int	10
Random seed (negative = random)	int	42

🏆 Search Results

id sklearn_search_results · ML / Tuning · Python export: yes

Every parameter combination a search tried, with its mean and spread of cross-validation score and rank — best first.

Inputs

Port	Type	Description
search	FittedModel	Output of a Hyper-parameter Search node.

Outputs

Port	Type	Description
results	DataFrame	One row per combination: parameters, mean_test_score, std_test_score, rank_test_score.