Section 3 kept leaving data on the table. country and segment are text, which no scikit-learn model accepts directly. avg_satisfaction has gaps that logistic regression refuses. And you’ve been scaling by hand, keeping track of X_train_s and X_test_s. Each of those steps has to be learned from training data only, applied identically to new data, and redone inside every cross-validation fold. Doing that by hand is how leaks and bugs happen. A pipeline does it for you.
A Pipeline is a list of steps: any number of transformers (objects with fit and transform, such as StandardScaler) followed by one final estimator. The pipeline behaves like a single model:
pipe.fit(X, y) fits the first transformer, transforms the data, passes it to the next step, and so on, then fits the model on the result.pipe.predict(X_new) runs X_new through the already-fitted transformers, then predicts.You’ve used make_pipeline, which names steps automatically ("standardscaler", "logisticregression"). Pipeline([("prep", ...), ("model", ...)]) lets you choose the names, which you’ll want next lesson when you tune parameters by name.
The real payoff is in cross-validation. cross_validate clones the entire pipeline and refits it in every fold, so the scaler’s means, the imputer’s medians and the encoder’s category list are learned from that fold’s training rows only.
Not every column needs the same preprocessing. ColumnTransformer routes each group of columns through its own transformer and glues the outputs side by side.
Three transformers do the work:
\nOneHotEncoder creates one 0/1 column per category: country_Egypt, country_Germany and so on. Set handle_unknown="ignore"; the default raises an error the first time a customer arrives from a country the training data didn’t include, which in production means a crashed prediction service.SimpleImputer fills missing values with the mean, median, most frequent value or a constant, learned from training data. With add_indicator=True it also adds a column that is 1 where the value was missing. Half of Cartwheel’s customers never answered the survey, and that fact may matter more than the score itself.StandardScaler, as before. A small make_pipeline lets one column group pass through two steps in order.import pandas as pd\nfrom sklearn.compose import ColumnTransformer\nfrom sklearn.impute import SimpleImputer\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.model_selection import StratifiedKFold, cross_val_score, train_test_split\nfrom sklearn.pipeline import Pipeline, make_pipeline\nfrom sklearn.preprocessing import OneHotEncoder, StandardScaler\n\nchurn = pd.read_csv(\"churn.csv\")\nX = churn.drop(columns=[\"customer_id\", \"churned\"])\ny = churn[\"churned\"]\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)\n\nnumeric = [\"tenure_days\", \"orders\", \"total_spent\", \"avg_order_value\",\n \"days_since_last_order\", \"used_coupon\", \"support_tickets\", \"mobile_share\"]\npreprocess = ColumnTransformer([\n (\"num\", StandardScaler(), numeric),\n (\"sat\", make_pipeline(SimpleImputer(strategy=\"median\", add_indicator=True), StandardScaler()),\n [\"avg_satisfaction\"]),\n (\"cat\", OneHotEncoder(handle_unknown=\"ignore\"), [\"country\", \"segment\"]),\n])\npipe = Pipeline([(\"prep\", preprocess), (\"model\", LogisticRegression())])\n\npipe.fit(X_train, y_train)\nnames = pipe.named_steps[\"prep\"].get_feature_names_out()\nprint(len(names), \"model inputs, e.g.\", list(names[8:12]))\n\ncv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)\nprint(\"all columns ROC AUC:\", cross_val_score(pipe, X_train, y_train, cv=cv, scoring=\"roc_auc\").mean().round(3))\nnumeric_only = make_pipeline(StandardScaler(), LogisticRegression())\nprint(\"numeric only ROC AUC:\", cross_val_score(numeric_only, X_train[numeric], y_train, cv=cv, scoring=\"roc_auc\").mean().round(3))The pipeline accepts the raw DataFrame, text and gaps included, and produces 21 model inputs: eight scaled numbers, the imputed satisfaction score and its missing flag, eight country columns and three segment columns. get_feature_names_out lists them with a prefix naming the transformer that made each one, which is how you match coefficients to columns later.
Then the surprise: the full pipeline ranks customers worse than the eight numeric columns alone, 0.776 against 0.807. Eleven one-hot columns and a satisfaction score with half its values imputed give logistic regression more coefficients to fit with only 298 rows, and on this data the extra columns carry more noise than signal. Egypt’s churn rate does look higher in the training data, but with 44 Egyptian customers in training that’s too thin to bet on.
\nThat’s not a failure of the pipeline. It’s the pipeline doing its job: it made testing the extra columns a five-line change, and cross-validation gave a clear answer. For now, the numeric model stays the champion. The categorical columns are worth revisiting when Cartwheel has a few thousand customers, or with stronger regularisation.
\n\n\nA pipeline also gives every setting an address: model__C is the C of the step named model, and prep__sat__simpleimputer__strategy reaches the imputer. Next lesson uses those addresses to search for the best settings automatically.
Pipeline chains transformers and a final model into one estimator: fit fits every step in order, predict runs new data through the same steps.","ColumnTransformer applies different preprocessing to different columns and concatenates the results.","OneHotEncoder(handle_unknown="ignore") turns categories into 0/1 columns and won’t crash on a category it never saw in training.","SimpleImputer(add_indicator=True) fills gaps and adds a column recording which values were missing.","Because the whole pipeline is refit in each fold, preprocessing only ever learns from training rows, which keeps cross-validation honest."],"further":[{"title":"Pipelines and composite estimators","url":"https://scikit-learn.org/stable/modules/compose.html"},{"title":"ColumnTransformer","url":"https://scikit-learn.org/stable/modules/generated/sklearn.compose.ColumnTransformer.html"},{"title":"OneHotEncoder","url":"https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html"},{"title":"Imputation of missing values","url":"https://scikit-learn.org/stable/modules/impute.html"}],"quiz":[{"q":"During cross_validate(pipe, X_train, y_train, cv=5), where does the StandardScaler inside pipe get its means from?
X_train, computed once before the folds are made.","why":"That’s what happens if you scale before cross-validating, and it’s the leak pipelines exist to prevent."},{"text":"From the validation fold, so validation data is centred correctly.","why":"Fitting anything on the validation fold would leak it into the model."},{"text":"From the test set, since that’s what will be predicted.","why":"The test set isn’t passed to cross_validate at all."}],"answer":0},{"q":"A customer from France (absent in training) reaches a pipeline whose OneHotEncoder uses the default handle_unknown. What happens?
handle_unknown="ignore", not the default."},{"text":"The encoder adds a new France column on the fly.","why":"The column layout is fixed at fit time; the model’s coefficients couldn’t use a new column anyway."},{"text":"transform raises ValueError: Found unknown categories ['France'].","why":"Correct. The default is "error". Use handle_unknown="ignore" for anything that will see live data."}],"answer":3},{"q":"Why use SimpleImputer(strategy="median", add_indicator=True) for avg_satisfaction?
Adding country and segment to the logistic pipeline lowers cross-validated ROC AUC from 0.807 to 0.776. What’s the best interpretation?
Make the starter pipeline handle gaps and surprises.
\npreprocess: scale the numeric columns; for avg_satisfaction, impute with the median and add a missing indicator, then scale; one-hot encode segment so that unseen categories are ignored instead of raising an error.pipe: steps named "prep" and "model" (a LogisticRegression), fitted on the training data. It should produce 13 model inputs.cv_auc: mean ROC AUC from cross_val_score with the provided cv.new_prob: the churn probability for new_customer, who comes from an unseen country, an unseen segment and has no satisfaction score.StandardScaler alone can’t handle NaN. Chain make_pipeline(SimpleImputer(strategy="median", add_indicator=True), StandardScaler()).","OneHotEncoder(handle_unknown="ignore"). The country column isn’t listed in the transformer, so it’s dropped.","Don’t forget pipe.fit(X_train, y_train); then pipe.predict_proba(new_customer)[0, 1]."],"explanation":"Eight scaled numeric columns, the imputed satisfaction score plus its missing indicator, and three segment columns make 13 inputs. Columns a ColumnTransformer doesn’t mention (here country) are dropped by default (remainder="drop"). The new customer gets all-zero segment columns and a median satisfaction with the indicator set to 1, so the model still produces a probability. Without handle_unknown="ignore" and the imputer, this one customer would crash a production service.