Structured and Unstructured Data

Structure is an agreement, not a value judgment

Structured data follows a fixed schema: database columns, spreadsheet cells, or sensor records with predictable fields. Unstructured data lacks that regular table shape: free text, images, audio, and video. Semi-structured data sits between them; a JSON log has named fields, but records may contain optional or nested shapes.

Models ultimately calculate with numbers, so unstructured inputs need a representation step. An image becomes a grid of pixel values; text may become token IDs for a language model or an embedding vector for retrieval. The original data is not inherently less organized - its useful structure simply is not already expressed as fixed table columns.

raw support email -> text cleanup -> numeric representation -> model
database record   -> validation   -> numeric columns       -> model

Scenario: A ticket classifier combines structured fields (priority, product) with unstructured message text. The product field is directly encoded as categories; the message needs a text representation. Both must still align to the same ticket row and label.

Tip: Preserve the raw source alongside transformed features when policy allows. A derived vector is useful to a model, but the original record is what lets a human audit whether preprocessing changed the meaning.