I was more thinking based on input data about things like what you are looking for in a partner, whether you are after a long-term relationship or not, etc. Then you throw in all that data to the ML deities and hope the output is something roughly related to what you're after.

Fair enough about word of mouth though, relationship durations are probably linearly correlated to the number of endorsements given by each customer - at some point in a relationship you end up with a rehearsed "so how did you two meet?"!

Actually I think this would be relatively easy, you're just looking for a bit of alignment in worldview and significant desperation.

The old ok cupid had a bunch of quiz questions that it'd match you for, both similarity and dissimilarity weighted by self reported importance.

Then Match Group bought them out and mostly gutted that system and turned it into another swiper :/

There's still eharmony, but I think that's just for white Christian boomers mainly.

The "limited" inputs include: All your conduct on the app, your "advertising" dossier.

Seems doable, frankly.

> I can't imagine there's a data scientist alive who could devise an algorithm that accurately, reliably predicts "kind of a good match but not really", given only some limited inputs (basic info + photos).

One approach would be to make the problem easier, just preferentially match people with a partner who is known to have short relationships and no/few awful reviews, based on history on the site.