In this case study

Dating

Deit orders its deck with a rating and a score. The rating is Elo, updated on every swipe. The score is a hand-set formula over six terms, recomputed daily. Neither is a trained model, but both decide what a person sees.

Elo

Every swipe updates the target's Elo with a K factor of 32, weighted by the swiper's rating. Three corrections are applied:

  • A converged replay retrains every Monday from the swipe history.
  • Pools are recentred per gender, so the two distributions are comparable.
  • A weak prior of five phantom pairs keeps never-liked profiles finite.

The power score

The deck is ordered by a power score, rescored daily:

Power score
power = 450·incomingLike
      + 300·exp(−(ΔElo / 400)²)
      + 250·0.5^(idleDays / 3)
      + 200·max(0, 1 − profileAgeDays / 7)
      + 100·completeness
      + 100·jitter

Reading the terms in order: a profile that has already liked the viewer comes first; profiles close to the viewer's own rating come next, with a bell-shaped falloff scaled to 400 Elo; recently active profiles beat idle ones, halving every three days; new profiles get a boost that decays over their first week; complete profiles beat sparse ones; and jitter varies the deck from day to day.

The daily pick

Deit dagsins, the daily pick, has a different objective from the deck. It optimises the chance that the viewer likes the card, rather than the probability of a mutual match. It requires the candidate's Elo to be above 1,050, and it never repeats a profile within a week.

Limits and schedules

The deck is refilled every six hours and top picks are computed at 11:00. Swipes are limited to 200 per hour and free likes to 20 per day. Dating is the one surface that exists only on mobile and only for users the identity model can confirm are 18.