It’s estimated that nearly 60 million Americans did at least some form of freelance work in the last year, many through the various platforms that have emerged to connect buyers and sellers. These platforms exist to make it easier to find the ideal match between customers and freelancers.

A problem with these platforms is that the platform itself doesn’t really know whether the freelancer will be good at completing any given task. There is also a significant degree of fluidity involved, with a high level of movement of freelancers as they come and go. This makes it difficult to ascertain the capacity of freelancers to perform each task.

Online matching

While in an ideal world the platforms would learn the preferences of each client through a process of trial and error, this wouldn’t be acceptable in the real world.

“The challenge is that you want to somehow very quickly learn the customer’s preferences based on the feedback or the outcome of the assignments,” the researchers explain.

It’s something known in the machine learning world as an exploration vs exploitation tradeoff. For instance, if the platform focuses on exploring new matches, then it could risk harming customer satisfaction. If they don’t explore, however, they may not be able to provide the best match.

The researchers developed an algorithm to make it easier to combine both approaches in an optimal way. The algorithm also takes into account the capacity of each freelancer and their uncertain availability.

“We formulate this as an optimization problem. There are some capacity constraints for every server and you have to make sure that you do not violate them,” the researchers explain. “In addition, every client is associated with a utility function of the received service rate and you need to maximize both the total utilities and the estimated matching payoffs.”

Testing for regret

The researchers put the algorithm to the test by measuring the so-called “regret rate”, which compares how the algorithm performs alongside an “oracle” that is able to know all of the dynamics and preferences of the client.

“We showed that the regret is very small and it decreases if you run the system for a longer time,” the researchers explain. “The regret also decreases if a particular customer assigns multiple tasks. In that case, the system gets increasingly good at learning the client’s preferences.”

The researchers hope that their work might help platforms to tackle the inherent uncertainty that exists and builds on previous work that has assumed a scenario where the arrival rates of different types of customers to the platform and matching payoffs were known beforehand.

“In our case, we don’t need to know this information,” they conclude. “We can dynamically allocate our assignments in response to these different arrival rates and matching payoffs. That’s the interesting thing about our algorithm and policy.”