Oversampling rebalances a dataset by increasing the number of minority-class examples — the simplest version just duplicates existing minority samples until the classes are closer to balanced.
Random Oversampling — The Basic Approach
Using the reference dataset (950 majority, 50 minority): random oversampling duplicates minority samples, with replacement, until the target balance is reached — e.g. duplicating the 50 minority examples up to 950, producing a balanced 950/950 dataset of 1900 total rows.
from imblearn.over_sampling import RandomOverSampler
from collections import Counter
ros = RandomOverSampler(random_state=42)
X_resampled, y_resampled = ros.fit_resample(X_train, y_train)
print(Counter(y_train)) # {0: 950, 1: 50}
print(Counter(y_resampled)) # {0: 950, 1: 950} -- balanced by duplicationThe Real Risk — Overfitting to Duplicated Points
Because oversampling literally repeats the same minority examples many times, a flexible model (especially KNN, decision trees, or neural networks) can start memorizing those exact repeated points rather than learning a general minority-class pattern — a distinctive overfitting risk unique to this technique. This is exactly the gap SMOTE addresses, by generating new, synthetic points instead of exact duplicates.
Applying It Correctly — Only on Training Data
from sklearn.model_selection import train_test_split
from imblearn.over_sampling import RandomOverSampler
# Split FIRST, on the original imbalanced data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, stratify=y, random_state=42)
# THEN oversample -- only the training set
ros = RandomOverSampler(random_state=42)
X_train_resampled, y_train_resampled = ros.fit_resample(X_train, y_train)
model.fit(X_train_resampled, y_train_resampled)
predictions = model.predict(X_test) # X_test stays untouched, naturally imbalancedThe test set must always reflect the real-world class distribution the model will actually face — oversampling it would give a distorted, unrealistic picture of production performance.
Practical Use Cases
- Smaller datasets, where undersampling's data-discarding tradeoff (see Undersampling) would throw away too much of the majority class
- A quick first technique to try before moving to more sophisticated approaches like SMOTE
Advantages
- Simple to implement and understand
- Doesn't discard any original data, unlike undersampling
Limitations
- Risk of overfitting to exactly duplicated minority points
- Increases dataset size and training time
Common Mistakes
- Oversampling before splitting into train/test — this leaks duplicated copies of the same underlying example into both sets, inflating test performance artificially.
- Always targeting perfect 50/50 balance by default — a more moderate ratio is sometimes a better fit, worth testing rather than assuming.
Interview Relevance
Q: "Why must oversampling happen after the train/test split, not before?" Oversampling before splitting can place duplicated copies of the same original minority example into both the training and test sets — the model then gets evaluated partly on examples nearly identical to ones it trained on, artificially inflating reported test performance.
Practice Question
You oversample the minority class to perfect 50/50 balance and see training accuracy jump to near 100%, but test recall barely improves. What might be happening?
