In the online dating context, an algorithm can get a good idea of my taste in partners by doing a similar comparison of me to other male users. Another male user of the site will have a similar taste in women to me if we are messaging the same women.
However, while this gives the algorithm a good idea of who I like, it leaves out the important factor of who likes me — my attractiveness to the female users of the site, measured by who is sending me messages.
Zhao’s crucial innovation is to combine information about both tastes and attractiveness. The algorithm keeps track of both who I am messaging, and who is messaging me. If a male user has similar taste (he is messaging the same women as I am) and attractiveness (he is messaged by the same women as I am) to me, we are scored as being very similar; if we are similar in one trait — if we have similar tastes but attract (or fail to attract) different groups of women, or vice versa — we have a moderate similarity ranking, and if we are different on both measures, we are counted as very dissimilar.
The researchers tested the model and found that it does a pretty good job. “[W]e illustrate that the new model performs well on recommending both unilateral and reciprocal contacts,” write Zhao, Xi Wang, Mo Yu, and Bo Gao. “In other words, the new model can better recommend partners that matches a user’s taste and attractiveness.”