Understand the idea
Training error measures fit to known examples. Raw accuracy can hide a rare class. A useful model needs evidence on held-away training folds, a relevant metric and a baseline. The simplest adequate model can be preferable to a complex one.
Python skill: Compare train_score and test_score arrays returned by cross_validate.
Meet the syntax
Pipeline([('prepare', preparation), ('model', estimator)])
cross_validate(candidate, X_train, y_train, cv=folds, scoring=metric)
cross_val_predict(candidate, X_train, y_train, cv=folds)
clone(candidate).fit(X_train, y_train)
validation['test_score']
f1_macro
balanced_accuracyPipeline- Keeps preparation with the model, so each fold learns fitted settings only from its training rows.
cross_validate- Fits fresh copies on training folds and returns their validation scores.
cross_val_predict- Makes one held-out training prediction per row for diagnosis; use cross_validate for the validation score.
clone(candidate).fit(X_train, y_train)- Starts an unfitted copy of the selected recipe and fits all development rows after selection.
test_score- Validation-fold scores inside cross_validate, despite the word test; the reserved final test is separate.
Follow the code
Use the numbered comments to connect each Python block to the workflow above.
from sklearn.model_selection import train_test_split, StratifiedKFold, cross_validate, cross_val_predict
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.dummy import DummyClassifier
from sklearn.base import clone
from sklearn.metrics import f1_score
# 1. Question, features available now, and later outcome
feature_names = ['queue_at_open', 'age_hours_at_open']
target = 'resolution'
X, y = df[feature_names], df[target]
# 2. Protect the final test; split X and y together
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.2, random_state=42, stratify=y)
# 3. Inspect training rows only
print(X_train.describe())
# 4. Define preparation and the candidate; no fitting yet
candidates = {'simple': Pipeline([('prepare', StandardScaler()), ('model', LogisticRegression(max_iter=1000))])}
folds = StratifiedKFold(n_splits=3, shuffle=True, random_state=42)
metric = 'f1_macro'
# 5. Compare a reference and candidates on matching training folds
reference = DummyClassifier(strategy="most_frequent")
reference_results = cross_validate(reference, X_train, y_train, cv=folds, scoring=metric)
cv_results = {'simple': cross_validate(candidates['simple'], X_train, y_train, cv=folds, scoring=metric, return_train_score=True)}
print('Reference validation:', reference_results['test_score'])
print('Candidate validation:', cv_results['simple']['test_score'])
print('Candidate minus dummy mean:', cv_results['simple']['test_score'].mean() - reference_results['test_score'].mean())
# 6. Assess the candidate with training evidence; inspect training-only errors
chosen_name = 'simple' # Sole candidate; report if it does not beat the dummy
diagnostic_predictions = cross_val_predict(candidates[chosen_name], X_train, y_train, cv=folds)
print(pd.crosstab(y_train, diagnostic_predictions))
# 7. Fit the fixed recipe; predict the final rows once
final_model = clone(candidates[chosen_name]).fit(X_train, y_train)
final_predictions = final_model.predict(X_test)
final_score = f1_score(y_test, final_predictions, average="macro")
print('Final metric:', final_score)
# 8. Explain the evidence and its limits in the self-review field
This practice: Run, predict and explain. Next: Change · When a good score is misleading.
The question and data dictionary
When a ticket enters the support queue, flag whether it will miss its service deadline.
One independent ticket per row in a stable staffing period. Late tickets are a minority; missing them matters. A class-balanced score is more useful than raw accuracy.
| Column | Meaning and availability |
|---|---|
| queue_at_open | Number of waiting tickets at entry. |
| age_hours_at_open | Hours since customer submission at queue entry. |
| resolution | Later outcome: late or on_time. |
| closed_late_flag | Flag entered at closure, derived from the outcome. |
The playground workflow
Use the same data boundary as ML → Workflow: frame, split, explore, prepare, validate against a reference, diagnose, then make the final evaluation. Each step answers one question.
- Question → X / yName the outcome and the inputs available when the prediction is needed. Later outcomes, identifiers and outcome-derived fields are not predictors.
- Split and protectReserve final rows before inspecting distributions or fitting. Independent rows permit a shuffled holdout; repeated entities or future forecasting need different designs.
- Explore training rowsInspect only development rows to identify types, missingness and class balance.
- Prepare → modelPut learned preparation inside a Pipeline so each fold learns it afresh. A numeric linear model can pass the original units through.
- Baseline → validateA dummy establishes what ignoring X achieves. Fit the candidate on matching training folds; these validation rows are not the final test.
- Choose → diagnoseChoose using training evidence, then inspect held-out training predictions. Simplify a model that fails to generalise; do not open the final test to choose.
- Final fit → predict → metricFix the recipe, fit all training rows, predict reserved rows once and calculate the declared metric from those saved predictions.
- Interpret and limitCompare validation with the baseline and final evidence. State units or class error costs, uncertainty from the small sample, and the population to which the claim applies.
Find the fault
Identify the failed decision, explain its consequence, then repair the workflow in the editor.
predictions = ['on_time'] * len(y_test)
# High accuracy means useful late-ticket detection?Explain your decisions · self-review
Before Run, write your prediction. After Run, explain what the evidence supports and what it cannot establish. Code checks cannot award these reasoning scores.
- Frame and boundaryName prediction time, observation, target and unavailable inputs. Choose a split that matches intended use.
- Preparation and baselineFit learned preparation inside each training fold. Compare the dummy on those same folds.
- Validation and selectionChoose from training-fold scores and error patterns. Explain any gap between training and validation.
- Metric and final testJustify the metric and interpret the reserved-test result. Do not revise the recipe using that result.
- Interpretation and limitsQuote baseline, validation and final results. Explain error units or costs, one failure case and one limit; avoid causal claims.
Use the interpretation field beside your output. After an attempt, Check identifies code evidence and shows reasoning guidance. The optional explained solution is one defensible approach, not the only acceptable answer.
Your inputs · SUPPORT120
120 synthetic independent observations. The dataframe df is supplied afresh on each Run. Column availability is described in the question’s dictionary above.
| queue_at_open | age_hours_at_open | resolution | closed_late_flag |
|---|---|---|---|
| 22 | 6.32038 | on_time | 0 |
| 20 | 0.733376 | on_time | 0 |
| 30 | 2.01174 | late | 1 |
| 23 | 2.62893 | late | 1 |
| 22 | 1.69297 | on_time | 0 |
| 19 | 7.67255 | on_time | 0 |
| 34 | 1.25077 | on_time | 0 |
| 9 | 0.933823 | on_time | 0 |
Your task · Follow
An always-on-time model can have high accuracy while finding no late tickets. Run the repaired macro-F1 workflow, inspect class-specific errors and the dummy, and explain which mistakes accuracy hid.
Required Python variables and evidence
Use these names so Check can inspect your workflow. Each meaning is shown beside its name.
| Variable | Meaning |
|---|---|
| target / feature_names | Your outcome column name and list of legitimate inputs. Choose from the data dictionary. |
| X / y / X_train / X_test / y_train / y_test | Original indexed feature/target data and aligned partitions. A 15–30% holdout is supported; choose and justify its seed/design. |
| folds / metric | A 3–5-fold shuffled KFold or StratifiedKFold object; metric is an sklearn scoring name. Regression: neg_root_mean_squared_error or neg_mean_absolute_error. Classification: f1_macro or balanced_accuracy. |
| reference / reference_results | Dummy estimator and its actual cross_validate result. |
| candidates / cv_results | Named Pipelines and matching cross_validate result dictionaries. Regression supports LinearRegression, Ridge and DecisionTreeRegressor; classification supports LogisticRegression, DecisionTreeClassifier, KNeighborsClassifier, GaussianNB, LinearDiscriminantAnalysis, SVC and MLPClassifier. |
| diagnostic_predictions | Actual cross_val_predict outputs for the chosen candidate on training folds; inspect residuals or a confusion table. |
| chosen_name / final_model | Name of your chosen validated candidate and a clone fitted on all training rows. Defend the choice, including any simplicity trade-off. |
| final_predictions / final_score | One saved final prediction array and its final metric. Positive error units for regression. No further selection after this call. |
Hint 1 — Think
An always-on-time model can have high accuracy while finding no late tickets. Run the repaired macro-F1 workflow, inspect class-specific errors and the dummy, and explain which mistakes accuracy hid. When a ticket enters the support queue, flag whether it will miss its service deadline. Which fields exist at that moment?
Hint 2 — Tools
Use a dataframe/series pair, train_test_split, Pipeline, cross_validate and a dummy suited to class labels.
Hint 3 — Approach
Keep final rows outside every fit. Compare matching training-fold evidence before predicting reserved rows in SUPPORT120.
Explained solution
from sklearn.model_selection import train_test_split, StratifiedKFold, cross_validate, cross_val_predict
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.dummy import DummyClassifier
from sklearn.base import clone
from sklearn.metrics import f1_score
# 1. Question, features available now, and later outcome
feature_names = ['queue_at_open', 'age_hours_at_open']
target = 'resolution'
X, y = df[feature_names], df[target]
# 2. Protect the final test; split X and y together
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.2, random_state=42, stratify=y)
# 3. Inspect training rows only
print(X_train.describe())
# 4. Define preparation and the candidate; no fitting yet
candidates = {'simple': Pipeline([('prepare', StandardScaler()), ('model', LogisticRegression(max_iter=1000))])}
folds = StratifiedKFold(n_splits=3, shuffle=True, random_state=42)
metric = 'f1_macro'
# 5. Compare a reference and candidates on matching training folds
reference = DummyClassifier(strategy="most_frequent")
reference_results = cross_validate(reference, X_train, y_train, cv=folds, scoring=metric)
cv_results = {'simple': cross_validate(candidates['simple'], X_train, y_train, cv=folds, scoring=metric, return_train_score=True)}
print('Reference validation:', reference_results['test_score'])
print('Candidate validation:', cv_results['simple']['test_score'])
print('Candidate minus dummy mean:', cv_results['simple']['test_score'].mean() - reference_results['test_score'].mean())
# 6. Assess the candidate with training evidence; inspect training-only errors
chosen_name = 'simple' # Sole candidate; report if it does not beat the dummy
diagnostic_predictions = cross_val_predict(candidates[chosen_name], X_train, y_train, cv=folds)
print(pd.crosstab(y_train, diagnostic_predictions))
# 7. Fit the fixed recipe; predict the final rows once
final_model = clone(candidates[chosen_name]).fit(X_train, y_train)
final_predictions = final_model.predict(X_test)
final_score = f1_score(y_test, final_predictions, average="macro")
print('Final metric:', final_score)
# 8. Explain the evidence and its limits in the self-review field
When a ticket enters the support queue, flag whether it will miss its service deadline. The reference ignores features. Training-fold evidence informs the choice, and the final metric describes only the reserved sample. Excluding later outcomes prevents answering the question with information unavailable at prediction time. The reasoning rubric needs human review.
Helpful prior knowledge: Repair a leaking workflow These links are guidance, not locks.
Sources and API context
Examples run with this Playground’s scikit-learn 1.4.2 / Pyodide 0.26.4 runtime.