Machine Learning · scikit-learn pipelines

Real tables mix numeric and categorical columns, each needing different preprocessing. A Pipeline chains the steps, and a ColumnTransformer routes columns to the right step:

• numeric: impute the median, then scale
• categorical: impute the most frequent value, then one-hot encode (handle_unknown="ignore" for categories never seen in training)

The whole thing is one estimator: fit, predict, cross-validation, grid search and saving with joblib all just work, and the leakage from last week can’t happen.

The data is a small DataFrame I generated in the snippet: synthetic student records with some values missing.

import numpy as np, pandas as pd, joblib, tempfile, os
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

rng = np.random.default_rng(0)
n = 500
df = pd.DataFrame({
    "attendance": rng.uniform(40, 100, n),
    "study_hours": rng.gamma(2, 3, n),
    "branch": rng.choice(["ICT", "CSE", "ME", "EE"], n),
    "hostel": rng.choice(["yes", "no"], n),
})
df["passed"] = ((df.attendance / 100 + df.study_hours / 15 + rng.normal(0, 0.25, n)) > 1.1).astype(int)
df.loc[rng.choice(n, 40), "attendance"] = np.nan
df.loc[rng.choice(n, 25), "branch"] = np.nan

num, cat = ["attendance", "study_hours"], ["branch", "hostel"]
pre = ColumnTransformer([
    ("num", Pipeline([("impute", SimpleImputer(strategy="median")), ("scale", StandardScaler())]), num),
    ("cat", Pipeline([("impute", SimpleImputer(strategy="most_frequent")), ("onehot", OneHotEncoder(handle_unknown="ignore"))]), cat),
])
model = Pipeline([("pre", pre), ("rf", RandomForestClassifier(200, random_state=0))])

X, y = df.drop(columns="passed"), df["passed"]
print("CV accuracy:", cross_val_score(model, X, y, cv=5).mean().round(3))
model.fit(X, y)
print("features after preprocessing:", list(model[:-1].get_feature_names_out()))

path = os.path.join(tempfile.mkdtemp(), "model.joblib")
joblib.dump(model, path)
print("reloaded predicts:", joblib.load(path).predict(X.head(3)).tolist())

Output:

CV accuracy: 0.768
features after preprocessing: ['num__attendance', 'num__study_hours', 'cat__branch_CSE', 'cat__branch_EE', 'cat__branch_ICT', 'cat__branch_ME', 'cat__hostel_no', 'cat__hostel_yes']
reloaded predicts: [1, 0, 1]