Stratified sampling divides data into groups and controls how much data comes from each group. It helps datasets reflect important subgroups rather than simply mirroring the largest population. section 6.3 section 6.4
Use stratification when dataset composition could distort model development:
- Balance customers from different markets, such as countries X and Y, across training and test datasets. section 6.3
- Reduce the influence of imbalanced or duplicate-heavy datasets by limiting data selected from defined groups. section 6.4
- Prioritize groups deliberately when equal representation is not appropriate, such as favoring fresh data over older temporal groups. section 6.4
Expanding from country X to country Y
A company that collected mostly customer data from country X must represent both X and newly entered country Y in its training and test datasets to pursue comparable service and model accuracy. section 6.3
