I’m talking about matching at the moment, because it’s a niche interest of mine and also because I attended a talk on mentoring. I’ve already spoken about how quickly matching people gets ugly and difficult. The programme leader was experiencing this first hand — having grown the programme by 500%, he was starting to struggle to make matches based on qualitative data. He asked me to write something up to offer potential solutions. Here’s my very brief rundown.
There are a few methods by which people are matched. These can be widely categorised into:
- Monogamous relationships
- Polyamorous relationships
- Relationships for the collective good
Let’s start with polyamory
Many-to-one, one-to-many, and many-to-many
This is the approach taken by most dating websites. The optimal outcome isn’t that one person is matched to one person. Instead the algorithm matches individuals to multiple others and then lets the users decide what to do next. Applying this approach to a mentor/mentee matching program is problematic, because the optimal outcome is a 1:1 exclusive pairing.
Relationships for the collective good
A typical example given for this kind of matching is employees and tasks. Every employee can carry out every task, but there’s a different cost in time — which I measure in dollars — for each employee. As a manager, I want to make sure I’m spending my employees’ time in the most efficient way.⁰ I lay out my table as in the previous blog and then find the “cheapest” route through.
Here the optimal matches are:
- Armond: Clean bathroom
- Francine: Sweep floors
- Herbert: Wash windows
As you can see, it would be easy to convert tasks to people and calculate a score based on a number of traits that you’d like to match. The reason I call these “collective good matches” is because the people in them have no say. Armond may prefer to wash windows because he likes to see the sky, for example, but it is inefficient at an organisational level to match him to that task — unless part of our matching score takes into account the employee’s desire to do the task.¹
Although this task is easy with a 3 x 3 grid, at scale it requires an algorithm to manage.
These relationships are a one-to-one match, like the ones above. We divide the population into groups and ask them to list their preferences. We can then apply an algorithm to match the two groups together. Unlike the above, where the group as a whole gets the best matches, this approach strives towards individual optimisation. This problem is usually stated as follows:
Given n men and n women, where each person has ranked all members of the opposite sex in order of preference, marry the men and women together such that there are no two people of opposite sex who would both rather have each other than their current partners. When there are no such pairs of people, the set of marriages is deemed stable.²
The population you’re matching is divided into two groups; proposers and reviewers. Proposers ask their highest preference; if the reviewer prefers this proposal to their current partner then they will dump their current parter and hitch themselves to the proposer; if not, they’ll turn down the proposer.³ Once every proposer has asked, the process begins again, continuing until the round in which no reviewers swap partners.
A consequence of this process is that the preferences of the proposers are always privileged over the preferences of the reviewer.
Applying a monogamous model to the problem of mentor-mentee matches benefits the preferences of one group over the other. If this approach were to be taken, I’d recommend making the mentor the group whose preferences are most respected, as they are the more scarce resource. By contrast, applying the collective good approach might lead to poorer individual matches but will ensure everyone gets a good match, as opposed to a mix of lower and higher scoring matches.
A hybrid approach that generates a “collective good” score and uses that to proxy the list of preferences that would be generated by a “monogamous relationship” approach seems to offer the best of both worlds. It would generate the preferences “blind” — avoiding gender and race bias that we all have — as well as removing the necessity for mentors and mentees to exhaustively research and analyse hundreds of profiles from the opposing group.
⁰ The manager in this scenario is of course entirely fictional
¹ Whether or not it is wise for managers to take their employees’ feelings into account when assigning tasks is not really a question: it is self-evidently a good idea.
² For the purposes of this problem, heteronormativity reigns supreme.
³ Equally, reviewers would rather have someone than no-one