What is data sampling?
Data sampling is how you choose which data to use out of everything available. The simplest form is random sampling, but the choice of strategy matters: stratified sampling preserves class proportions, balanced sampling counters class imbalance, and targeted strategies pick the most informative examples. Each choice biases what the model sees and how representative your evaluation is.
Getting sampling right is often the difference between a fair, useful dataset and a skewed one.
Key takeaways
- Sampling selects a subset of data for training, evaluation, or labeling.
- The strategy, random, stratified, balanced, or targeted, shapes the outcome.
- Poor sampling introduces bias and unrepresentative evaluation.
How it works
For a representative split you might sample randomly or stratify by class so proportions match the source. To address imbalance you might oversample rare classes or undersample common ones. To spend labeling budget well you move toward informed strategies like coreset selection and uncertainty sampling, which choose by diversity or model confidence rather than chance.
Why it matters
Sampling decisions quietly determine what a model learns and whether its evaluation reflects reality. Thoughtful sampling counters imbalance, builds trustworthy test sets, and directs annotation toward the most valuable examples, all of which are central curation concerns.
Frequently asked questions
What is stratified sampling?
Sampling that preserves the proportion of each class or group, so the subset mirrors the composition of the full dataset.
How does data sampling relate to active learning?
Active learning is a targeted sampling strategy that repeatedly selects the most informative samples to label using model feedback.
Related terms