Dating App Matching Algorithms, Explained Without the Hype

Start with rules, not models

Every successful matching system we have shipped started as a rules engine: distance, age range, orientation, activity recency. That gets a new app to a usable feed on launch day, and it produces the interaction data any model will later need.

A machine learning ranker with no behavioural data is worse than a sensible sort by recency. Sequence matters.

Desirability scoring and its trade-offs

Elo-style desirability scores rank profiles by how often they are liked by people who are themselves selectively liked. They work — they also concentrate attention on a small pool and make the experience hostile for everyone else.

The counterweight is exposure balancing: cap how often any profile appears per day, and boost under-served accounts. Retention is a two-sided problem.

Reciprocity is the metric that matters

Optimize for mutual likes that turn into replies within 24 hours, not raw swipe volume. Reciprocal recommendation ranks candidates by the probability that both parties act, which is a different objective from ordinary recommendation and needs its own training labels.

Practical architecture

A candidate generator narrows millions of profiles to a few hundred using cheap geo and preference filters. A ranker then scores that shortlist. Keeping generation and ranking separate lets you upgrade the model without touching the feed infrastructure.