preference_optimization

Given a top N list of preferences between lists A & B and X maximum matches, maximise...something.

Download

This is a typescript package, but should be interoperable with Javascript. Check out this package on npmjs.com

Example of Use

const input = {
  clients: {
    clientA: ["companyA", "companyB"],
    clientB: ["companyB", "companyC"],
    clientC: ["companyC", "companyA"]
  },
  companies: {
    companyA: ["clientA", "clientB"],
    companyB: ["clientA", "clientB"],
    companyC: ["clientA", "clientB"]
  }
};

const meetings = 2;

const output = buildScheduleFromScores(
  input.companies,
  input.clients,
  meetings
);

const outputExpectedToBeSimilarTo = {
  schedule: [
    { companyA: "clientA", companyB: "clientB", companyC: "clientC" },
    { companyB: "clientA", companyC: "clientB", companyA: "clientC" }
  ],
  matching_score_totals: {
    facilitators: { companyA: 5, companyB: 6, companyC: 4 },
    participants: { clientA: 7, clientB: 5, clientC: 3 }
  },
  participant_schedules: {
    clientA: ["companyA", "companyB"],
    clientB: ["companyB", "companyC"],
    clientC: ["companyC", "companyA"]
  },
  facilitator_schedules: {
    companyA: ["clientA", "clientC"],
    companyB: ["clientB", "clientA"],
    companyC: ["clientC", "clientB"]
  }
};

"Expected to Be Similar To"? That Looks Like an Impure Function!!!

Yes, this library uses a MonteCarlo shuffling system to decide which matches get priority ("first dibs") with each schedule timeslot that gets filled with matches.

So what does it optimise for? And How?

See below for 'what' the optimisation process seeks to find, but specifically, the algorithm generates 20 random outcomes and picks the best one.

Explanation of Problem

This problem came about when I was involved in a 'Speed Dating-esque' recruiting event.

One would read about the companies beforehand and make a ranked preference order choice, the companies would do the same for the candidates. Then a schedule was drawn up for 8 possible sessions.

Apparently this process was done manually in a very tedious Excel process, so I decided to build a system that would automatically produce an optimum result.

Naturally, the 'optimum result' may be different depending on your point of view, but one would assume that the following would apply:

1. 📅 All possible sessions should be filled, even if the matching is not optimal (in this example: the companies are guests, so should therefore see as many job-seekers as they can)
2. 🗳️ Ranked order preference is respected in that a weight can be given to the ranking, and perhaps for mutual matches.

I'm hoping to build a somewhat generalised system that allows the tweaking of scoring biases and other parameters so that this can act as a library for similar problems.

Code

The code is written in Typescript.
I intend for the outside of the function to not show this, but internally the terms Dogs🐶 and Cats🐱 are used quite a lot.

• Dogs refers to the Companies in the Problem Example, the main schedule and scores are generated from the perspectives of the dogs, when building the schedule, dogs will always be given a full schedule (if possible)
• Cats refers to the Job-seekers in the Problem Example. I intend for them to be given a separate 'cat-specific' version of the schedule, but as a result of the scheduling, cats may end up with empty slots in their schedules 🙀

I used these terms because they're shorter, more fun and a little less hard to read than Companies and Job-seekers or Clients 😹

Running Tests Locally

Install Dependancies with yarn Run Tests with yarn test