Algorithmic mechanics

Algorithms curate who sees whom, in what order, how often, under what conditions. They are not neutral intermediaries.

From ELO to dynamic engagement

Early Tinder used a chess-style ELO-like rating: every profile had a hidden attractiveness score; a right-swipe from a high-rated profile raised yours much more than one from a low-rated profile. [21] The system produced rigid hierarchies and “condensation” — a small elite cycled algorithmically while the majority sank into invisibility.

Modern apps have officially moved off pure ELO. Current models weight dynamic engagement, location, real-time activity and selectivity:

• A user who right-swipes everything (rate → 100 %) produces not a match signal but a spam signal.
• The platform then reduces visibility (shadow-ban-like effects).
• The desperation-driven shotgun strategy algorithmically sabotages one’s own potential.

Gale-Shapley and the stable marriage problem

A more theoretically rigorous approach, used in Hinge’s “Most Compatible” premium feature, is based on the Nobel-prize-winning Gale-Shapley algorithm. [22] It seeks stable matchings in two-sided markets: no two people should prefer each other over their currently assigned partner.

Empirically the algorithm internalizes gendered trade-offs: holding looks roughly equal, women weight socio-economic signals (income, education) more strongly. [22] Online sorting still differs from real-world marriages — offline “search frictions” matter in ways data-saturated online space drops.

Collaborative filtering

Most recommendation logic is dominated by collaborative filtering: users with similar swipe patterns get similar recommendations. In the dating context, this is broken because:

1. It is one-sided — recommendations ignore whether the other person would like back.
2. It exponentially amplifies majority preferences (popularity bias, filter bubbles). [26]
3. Marginalized and non-normative profiles slide into invisibility. [27]

Reciprocal Recommender Systems (RRS)

Because a match requires mutual consent, current research treats reciprocal recommender systems: a score not for one side but for the probability of mutual interest. [28, 29, 30]

A typical setup:

• Interest similarity (Jaccard over people I contacted).
• Attractiveness similarity (Jaccard over people who contacted me).
• Harmonic mean instead of arithmetic — prevents extreme imbalances from inflating the score.

Frameworks like FAIR-MATCH combine reciprocal scoring with fairness constraints (demographic balance). [26] Accuracy in benchmarks: collaborative filtering ≈ 25.1 %; reciprocal methods ≈ 28.7 %. [26]

What that means

• Like-spraying is no shortcut — the algorithm recognizes it.
• Reciprocity is the lever modern apps actually push.
• Premium-tier recommendations are not magically better; they use the same dataset with wider reach and “uncovered scarce profiles” (see monetization).

Visibility mechanics are a consequence of technical design choices, not laws of nature. Knowing them does not let you trick them — it lets you name them.