ML Pipeline Simulator

In the Data block, press Your data…. There are two ways in.

Once loaded, the simulator lists each column with its type (number or category), how many values are missing and how many distinct values it has. Pick the target column, the task (classification or regression) and, in a binary classification, the positive class. Untick the columns that should stay out (identifiers, free text) and click the type of a numeric column to treat it as a category (like Pclass in the Titanic data).

The simulator always builds the same scikit-learn skeleton, the one from the preprocessing and pipelines lessons:

prep = ColumnTransformer([
    ("num", Pipeline([("impute", ...), ("transform", ...), ("scale", ...)]), num_cols),
    ("cat", Pipeline([("impute", SimpleImputer(strategy="most_frequent")), ("encode", ...)]), cat_cols),
])
pipe = Pipeline([("prep", prep), ("reduce", ...), ("model", ...)])

• Per column (in the Preparation node) lets a column leave its lane, with its own imputation, transform and scaling (or another encoding, if it is categorical). Columns with the same choices become one branch of the ColumnTransformer, such as ("num_log1p", Pipeline([...]), num_log1p_cols) in the exported code.
• The Explore tab shows the training rows (X_train) before any model: the target, each column (histogram by class, skewness, outliers, missing values), the association between columns and suggestions with buttons that change the pipeline. Each pair of columns is measured the way its types allow: Pearson's r for number × number, the correlation ratio η for category × number and Cramér's V (bias-corrected) for category × category; clicking a cell draws the pair as a scatter, one box per category or the make-up of each category. The test rows stay out here too; whatever you decide by looking at the data has not peeked at them.
• The train/test split comes first (train_test_split, stratified in a classification). Everything is fitted on the training rows only: the imputer's mean, the scaling, PCA, feature selection and the model.
• Cross-validation runs on the training rows; the test set is held out and evaluated once. The metrics table shows the three columns side by side: training far above validation is overfitting.
• Permutation importance shuffles one original column at a time in the test rows and measures how much the main metric drops.
• The Fit the preprocessing on ALL rows option is there for you to watch leakage happen: with it on, the preprocessing sees the test rows before the evaluation, and the numbers become optimistic.
• Compare models puts several models through the same preparation and the same folds (cross_val_score for each) and shows every fold's score; Use puts the chosen one in the pipeline.
• Grid search runs a GridSearchCV over one or two hyperparameters (up to 100 combinations): with one, it draws the validation curve; with two, a heat map. The best combination is scored on the test set once, and Apply to the pipeline copies the values into the model.
• The Code button gives, in Python, what the open tab shows: the last run, the comparison or the grid search.
• Download the prepared data (in the Preparation block) and Download the reduced data (in the Reduction block) save a .csv with every row as the model receives it: the steps are fitted on the training rows, the split column marks train and test and the last column is the target.
• The 2D map tab draws the training rows in two dimensions with t-SNE (scikit-learn), UMAP (umap-js, because umap-learn needs numba, which Pyodide lacks) or PCA's first two components, after Preparation or after Reduction too, colored by the target. It is only for looking: nothing there enters the pipeline. Above 50 columns, t-SNE and UMAP first get 50 principal components.
• The browser runs scikit-learn 1.8. Colab may run another version, so the last decimals can differ.

Your pipeline is saved in this browser for each scenario; ↺ Reset brings the original back.