Workflow & Validation

Stabilization

Definition

The point in a continuous learning workflow where more training stops meaningfully changing the model's rankings.

Why it matters

It is a common, though not the only, signal that review can wind down.

In practice

Stabilization is the point where more training stops changing the model much. In TAR 1.0 it is usually judged from the control set: precision and recall at the chosen cutoff stop improving from one training round to the next.

In continuous active learning the comparable signal is a flattening gain curve: each new batch adds few responsive documents compared with earlier batches.

Stabilization says the model has learned what it can from the training so far. It is not a measure of how many responsive documents were found, so it is normally followed by validation.

Compare with

  • Stopping Criteria: Stabilization is one possible input to a stopping decision; the stopping criteria are the agreed rules for ending review.
← Back to glossary
Stabilization Definition - TAR Glossary | eDiscovery Decoder