Data science and machine learning · BEGINNER LESSON
Imbalanced dataset
An imbalanced dataset is a data table where one class has many more examples than another class. In this setting, what is measured is the number of samples in each class, and sometimes how model accuracy changes for rare cases.
What it is and what is measured
An imbalanced dataset has unequal class counts: one class is the minority class and another is the majority class. The imbalance is measured by comparing how many samples belong to each class, not by the labels themselves. In real examples, rare events such as collisions can create this pattern. The practical concern is that a model may be less accurate for the underrepresented class, even when overall accuracy looks good. [0,1,2]
[1] Assessing data imbalance correction methods and gaze ... [2] Kernel-based dynamic ensemble approach for classifying ... - PMC [3] Enhanced skin cancer classification for minority classes using ... - PMC
Why imbalance changes model behavior
When one class dominates the training data, a model can learn patterns that favor the majority class. That can make it predict rare cases as if they were common, which lowers performance on the minority class. Some studies also note that overlap between classes in feature space, or the space of measured variables, can make separation harder. [0,1,2]
[1] Assessing data imbalance correction methods and gaze ... [2] Kernel-based dynamic ensemble approach for classifying ... - PMC [3] Enhanced skin cancer classification for minority classes using ... - PMC
Common ways people try to fix it
A common data-level approach is resampling. Under-sampling removes examples from the majority class; over-sampling repeats minority-class examples; SMOTE, the Synthetic Minority Over-Sampling Technique, makes synthetic minority samples by interpolating between a sample and its k-nearest neighbors, where nearest neighbors are the most similar points by distance. Some studies also use class weighting, normalization, or ensemble methods. [0,1,2]
[1] Assessing data imbalance correction methods and gaze ... [2] Kernel-based dynamic ensemble approach for classifying ... - PMC [3] Enhanced skin cancer classification for minority classes using ... - PMC
Limits and cautions
These fixes do not always improve results. Under-sampling can leave too little data, and simple over-sampling can duplicate examples and increase overfitting, which means a model may memorize training data instead of learning general patterns. One study found that rebalancing could reduce accuracy, worsen calibration, and change the data distribution away from what occurs in the real world. Another warning is that a small dataset may be incomplete rather than truly imbalanced. [0,1,2]
[1] Assessing data imbalance correction methods and gaze ... [2] Kernel-based dynamic ensemble approach for classifying ... - PMC [3] Enhanced skin cancer classification for minority classes using ... - PMC
SEE THE IDEA
How class imbalance affects modeling
One class has many more examples than another.
The model sees the majority class more often and may learn biased rules.
People may remove majority examples, repeat minority examples, create synthetic ones, or change class weights.
Performance should be checked on rare cases, not only overall accuracy.
Unequal counts influence training, and correction methods can help or harm depending on the dataset.
Imbalance changes what the model learns; fixing it requires careful evaluation.
TRY IT YOURSELF
Test your understanding
Choose an answer, then check the explanation.
PUT IT TOGETHER
Match each concept to its meaning
Drag a label onto its description, or select a label and then select a description. This also works with a keyboard.
4 pairs to match
Sources and further reading
This general lesson was generated from public source material. It explains a concept; it does not interpret a private paper or provide medical advice.