Why Online Matchmaking Could Be 45 Per Cent Better

Finding a partner through online dating websites is hugely popular but notoriously hit-and-miss. Now one research team has worked out how to improve the process by studying the behaviour of people on these websites 

Online dating has never been bigger. About one in ten Americans have used an online dating website and of these 66 per cent have gone on to date somebody they met online, according to the Pew Research Internet Project. That’s big business, generating over $1 billion per year in revenue.

So finding better ways to match couples could be worth big bucks. Today, Kun Tu at the University of Massachusetts in Amherst and a few pals say they’ve worked out how. Their insight comes from mining the data describing the behaviour of people who sign up. And they say their new technique improves the rate of successful matches by as much as 45 per cent.

Dating websites often work by asking suitors to describe their ideal partners in terms of parameters such as age, height, income range, marital status, number of children and so on. People can search for others or have them recommended based on these choices.

This process can be uncomfortably asymmetric. Although people might send messages to those who match their preferences they may not necessarily receive them from these ideal matches. “Balancing the expectations of the initiator and the receiver is a challenging task,” say Tu and co.

The new technique to match people using both parties’ preferences. But there is also a twist—it turns out that people’s actual preferences differ substantially from their stated ones.

To gain their insight, Tu and co guys studied the behaviour of 2 million male-female pairs on the Chinese dating website Baihe.com over three months. Their dataset consisted of all the metadata associated with messages sent between these pairs—the date and whether it received a reply as well as each party’s profile which includes 21 parameters. These are things like gender, age, body type, education, occupation, annual income level, Western/Chinese zodiac sign, number of profile photos, city of residence and details of any children.

When it comes to replying to messages, it turns out that people are not as fussy as they think. When faced with messages from a selection of potential partners, people often relax their requirements and reply to those who do not match their stated preferences.

So Tu and co developed an algorithm that learns each suitor’s actual preferences (as opposed to their stated preferences) by analysing the types of people they respond to. And it turns out that the most important parameters boil down to age, has/lives with children, weight, income and height.

This algorithm then goes on to recommend other potential suitors who match each other’s actual preferences.

To test this algorithm, Tu and co examined each suitor who had sent more messages to more than one person. Then then used their algorithm to analyse these receivers and decide who was most likely to reply. They then deleted the 50 per cent of the messages that were least likely to succeed.

They repeated the same process using the conventional matchmaking procedure, deleting half the messages according to how well the receivers matched only the stated preferences.

The results are an eye-opener. Tu and co say that messages routed by their algorithm are 45 per cent more likely to receive a response than messages routed according to stated preferences.

That’s a significant difference that could be worth a few bucks in a $1billion market. Of course, there are challenges ahead. One interesting questions is whether an algorithm that works in China also works in the US or Europe. Another is whether it works as well in the real world in real time as it does on historical data.

There’s only one way to find out, of course. Perhaps Tu and co are looking for their ideal dating website partner right now.

Ref:arxiv.org/abs/1401.8042 : Online Dating Recommendations: Matching Markets and Learning Preferences