What is a generalization hierarchy, and what happens to privacy and utility as you climb it?

A generalization hierarchy defines progressively broader versions of a value; moving up increases privacy but destroys utility.

* A generalization ladder: climbing trades utility for privacy; algorithms climb only as high as needed. *

Each quasi-identifier gets a ladder of abstraction, e.g. for Age:

• L0 — Exact values (28): maximum utility, zero anonymization.
• L1 — Small ranges ([25–30]): slight precision loss.
• L2 — Wider ranges ([20–30]): meaning starts to blur.
• L3 — Very broad ([0–20], [20–40]): significant utility loss.
• L4 — Full suppression (*): maximum privacy, zero utility from that attribute.

The anonymization algorithm searches for the minimum generalization level that still hits the target k — climbing only as high as necessary, to preserve as much utility as possible.

Tip: Privacy and utility sit on a seesaw. Good hierarchies (and good algorithms) let you buy the privacy you need with the smallest possible utility payment.

Go deeper:

• k-anonymity (Wikipedia) — generalization ladders and the privacy/utility trade.
• k-Anonymity (Sweeney, 2002) — formal treatment of generalization hierarchies.