Volatility

Whether an attribute has ever been replaced — read from the supersession history rather than asserted about a category.

Why it matters in a memory layer

It is free. The record already knows which beliefs were retired, so no judgement is needed about which kinds of fact "tend to be durable" — which is fortunate, because that judgement gets it wrong. On this course's corpus diet and residence both sound stable and both have changed; five of six attributes have already been replaced once.

The one attribute reported stable is stable because the user moved house once and has not moved again. So stable means not yet observed to change, and reporting it as a property of the attribute overclaims from an observation window of one.

Count supersessions for the subject you are modelling. The slot with the most churn on this corpus belongs to the user's partner, and a volatility number computed over the wrong subject is worse than none.

Connections

Taught in: From Facts to a User Model

Do not confuse with: Salience