Metric

Metric in SuperMemo is a measure of the accuracy of the spaced repetition algorithm, or one of its sub-components.

Different metrics may be used in different contexts. Of particular interest is a universal algorithmic metric for cross-comparisons between spaced repetition algorithms.

If we have a set of recorded repetition scores (wins): W₁, W₂, W₃, …, W_n, we would like to measure the accuracy of retrievability predictions by a spaced repetition algorithm. The predictions (bets) that we want to assess are denoted as: B₁, B₂, B₃, …, B_n. The metric that would measure the accuracy of recall predictions is called a recall metric.

The value of a recall metric is to assess spaced repetition algorithms without the need to run an actual learning process (which may take decades). With a single set of repetition histories (set of wins), we can simulate the execution of any imaginable algorithm and generate the set of predictions (bets). An example of a useful recall metric is the B-W metric.

An algorithmic metric is a metric than can compare two individual recall metrics for two different spaced repetition algorithms. The universal algorithmic metric should, ideally, be reflexive, symmetric, transitive, and equal to zero for a perfectly predicting algorithm.

See: Universal metric for cross-comparison of spaced repetition algorithms

This glossary entry is used to explain SuperMemo, a pioneer of spaced repetition software since 1987