Class weights versus resampling in a controlled comparison

Class weights change how much each training error contributes to the objective. Random oversampling repeats minority rows so they appear more often during fitting. These operations can produce similar decision boundaries for some models, but they also change the raw probability scale and interact with regularization.

Hold the model and features constant

Our training period contains 2,795 negatives and 205 positives. We fit the same standardized logistic regression three ways:

• unweighted on the original rows;
• class_weight="balanced" on the original rows; and
• randomly oversampled to 2,795 rows per class.

The seed, three numeric features, regularization and validation set stay fixed.

Ranking is similar in this case. The weighted and oversampled probability means are far above validation prevalence, 7.53%, and their default-threshold workloads are radically larger. That does not make weighting inherently wrong. It shows that a rebalanced fitting objective does not automatically yield probabilities for the original population.

The LogisticRegression documentation defines class weights and notes their multiplication with sample weights. The chosen regularization convention matters when duplicated rows change the effective objective.

from imbalance_cases import weights_case

r = weights_case()
assert r['training_counts'] == {0: 2795, 1: 205}
assert r['resampled_counts'] == {0: 2795, 1: 2795}
v = r['validation']
assert v['class_weight_balanced']['selected_at_half'] > 400
assert v['unweighted']['selected_at_half'] == 9
print(r)

Run the comparison in the imbalanced-model lab. The result is scoped to one seeded synthetic split and one logistic specification.

Compare what the intervention actually changes

Weights are economical because they do not enlarge the dataset. They can express asymmetric importance directly and work well with estimators that support sample weights. Oversampling can help batch-based learners see minority examples more often, but exact duplication adds no new feature patterns and may encourage memorization.

SMOTE adds interpolated patterns rather than duplicates, introducing a separate neighbourhood assumption. Undersampling reduces majority computation but discards observations. None of these choices replaces feature-quality work or a suitable split.

Evaluate ranking on original validation prevalence. If probabilities drive cost, capacity or expected-value decisions, calibrate with a disjoint set drawn from the intended population. Select thresholds after the probability model is finalized. Report both the development intervention and the population on which metrics were computed.

Exercise: vary logistic C across a predeclared grid for all three strategies inside cross-validation. Compare coefficients, average precision, Brier score and selected workload. Explain why changing C can prevent the weighted and duplicated objectives from being numerically equivalent.

NeuraPath's Data Science course treats rebalancing as a controlled intervention. A convincing comparison holds the estimator fixed, keeps validation prevalence natural, and checks probability behaviour as well as ranking.

Continue learning

This article is part of the Imbalance, calibration and decision thresholds sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in SMOTE inside cross-validation without leakage.
• Continue with Calibrate predicted probabilities on held-out data.