Datasets

A model is only as defined as the rows it is trained on. After this chapter the reader can name feature types, find a dataset, and judge whether the split and the label balance are fit to use.

The first question about a row is what each column is, and Feature Types separates a label from an ordered rank from a real number, because encoding and regression treat those differently. Knowing the types is useless if nobody can find the data, so Data Catalogs is the shelf that names what exists and the tools that implement one. A found dataset can still look good in training while its labels are noisy or its easy examples dominate, and Dataset Confidence uses dataset cartography to decide how much to trust it. With trust settled, Datasets is the work of typing the data, splitting it, and deciding when the sample is fair, from structured versus unstructured data through imbalance and transfer learning.