Sampling selects a useful subset of observations when a dataset is too large, imbalanced, or repetitive to use efficiently. section 6.4
Use sampling when:
- A defined process generates large amounts of data without manual labeling, such as clickstreams from search engines or marketplaces. section 6.4
- Duplicates or severe imbalance make indiscriminate training inefficient. section 6.4
- Stratification can divide observations by key features or clusters, while assigning more weight to relevant groups such as recent data. section 6.4
A practical sampling decision
For a large clickstream dataset, split records into meaningful groups, limit redundant observations, and prioritize recent groups when time affects the problem. section 6.4
