Objectives¶
The primary objective is binary cross-entropy for the observed separation hazard. Match quality enters the hazard with a positive constrained coefficient, so a higher score reduces the predicted separation probability conditional on the configured baseline.
Two optional objectives can be enabled alongside the hazard loss:
| Objective | Rows used | Purpose |
|---|---|---|
| Wage growth | Periods with an observed future wage target | Predict masked future wage growth with MSE |
| Destination | Separations with a known next employer | Rank the observed destination against sampled firms |
The destination loss excludes the incumbent firm from candidates. Evaluation destination recall, MRR, and NDCG are consequently defined only for observed moves with a known next employer.
The precise weights and enabled heads are configured under model and
training; see Configuration.