Algorithmic mechanics

Recommendation systems make several decisions

Candidate retrieval, ranking, activity and estimates of mutual interest are separate tasks. A familiar algorithm name therefore cannot explain an entire app.

As of September 1, 2026, Tinder says it no longer uses Elo. It describes activity, distance, preferences, profile information and likes/dislikes among its inputs. This is a provider account, not independent disclosure of every weight. 9

Hinge describes a deep-learning recommendation system introduced in 2025 that considers mutual interest. Describing its current system simply as Gale–Shapley is insufficiently supported. 10

The laboratory’s terms

Elo illustrates changing ratings. Gale–Shapley refers to stable assignment, but our variant adds selection and rejection rules. Collaborative filtering learns from interaction patterns; our simplified engine does not reconstruct a production system.

Reciprocal and Reciprocal without interventions are our own teaching models, not validated performance improvements. Both directions currently use the same interest similarity, which substantially limits the interpretation.

Source correction

The previously cited FAIR-MATCH paper was withdrawn on January 28, 2026 because of critical errors. Its performance claims have been removed; the withdrawal is retained as correction history. 11

Reciprocal recommendation research exists independently. Kawamura and colleagues’ 2025 work studies exposure-bias correction using DBLP and synthetic data. It does not establish improved Hinge date or match rates. 12