Data Science & Online Dating

It’s one of the most notable features of e-commerce websites like Amazon: the recommendation engine. “Based on your purchasing history you might also enjoy the following books”. Data Science has been fuelling the advancement of these recommendation engines, increasing conversions to order. But while matching people to products is one thing, matching people to people is a different kind of endeavour – how successfully is data science informing online dating?

The popularity of online dating

Human relations are probably the biggest existing market. At some point of time, almost everyone seeks to find a partner for life. For these encounters, social circles, school, university, work or leisure activities are the most probable filters. While each of these layers naturally ensures some sort of compatibility (common interest / background), initial uncertainty remains and the two individuals will have to invest significant time to discover their fit. Enter online dating: The rise of online dating platforms has not only significantly increased the accessible pool of “potential partners”, but also promises to limit the necessary upfront investment (primarily time) by increasing the quality of partner recommendation through smart matching algorithms.

The matching algorithms

The specifics of the matching algorithms are highly proprietary (with the notable exception of OKcupid), but essentially they always work along the following lines:

  • Taking the user’s preferences (taken from a questionnaire) and basic information and transforming this into character dimensions
  • Using these multidimensional user profiles to calculate core compatibility or affinity scores between users

Online dating platform are operating within an interesting dilemma: while every user should be engaged (through high quality and frequency of matches), the companies also have the incentive to maintain the user on the platform as long a possible (more months of active payment). Nevertheless, better matching and recommendation quality across the platform on average is essential for business success and companies like to showcase their efforts in this regard.

Problem of the target variable: When is a recommendation a success?

However, to really assess the quality of the recommendation or matching algorithms, it is essential to know what to benchmark against. While “initiation of communication” might be a first indicator, the real quality of the matching algorithm only reveals itself at later stages – in real life.

Interestingly, when average customer lifetime decreases, while communication between users on the platform increases, online dating companies could theoretically attribute their matching algorithms some success. The target variable – a successful relationship – remains tough to measure.

So for now, conveying the impression of a better “filter” than randomness in real life and increasing communication on the platform might be the only indicators to track reliably. The real quality of matching algorithms might remain in the shadows for much longer. – the patch towards full automation?

The new technicality of online dating has however also enabled a lot of interesting projects, which underline the absurdity of the science of matching by giving it an additional twist of automation. Justin Long, a computer scientist from Canada, has recently programmed the small social bot (some might call it artificial intelligence), which learns the user’s preferences regarding profile pictures and even draws on natural language processing to streamline some of the online conversation, which usually revolves around initiating contact.

But even with those advancements, the real litmus test for compatibility will always be conducted in real life. And any data science empowered attempt towards the improvement of matching quality should be cherished. Even though its assessment remains as tough as dating as such.