by Martin Rusis | 12 months ago
The topic of Machine Learning in matching platforms is still new to many. However, getting a grip on its fundamentals is actually pretty straightforward. This article outlines the main ideas, concepts and surprises that we at CNXION often cover in briefing clients about how Machine Learning applies for their marketplace or matching platform concept.
While both search and matching connect user desires with provider offers, the method they go about doing this is categorically different. Moreover, the offers made by matching enabled by Machine Learning can be orders of magnitude more relevant and effective for the demander than conventional search results.
In a conventional search, the active user seeks a passive item or ‘asset’. The access to this is governed by a reactive rule-based system. The impetus and responsivity is all on the user, or ‘requester’, side.
Matching, however, has active users, active providers and active ‘assets’. While the user still does their part, the provider is likewise trying to find them and, most important, the system itself learns how to facilitate that for both sides most efficiently.
In the past, it took a lot of money to connect users and providers profitably. Now, you can use targeted Machine Learning to dramatically cut the overheads.
The key is that Machine Learning works best with large amounts of accurate data applied to defined qualitative systems. Therefore, small marketplaces that are highly targeted can sharply define their ‘system’ and deliver results at the same or better accuracy than larger and less-focused marketplaces.
The matching takes place in how the connection between demander and provider is made. This differs from traditional marketplaces where a match occurs when a requester query is processed centrally and filtered through a larger library system. In some ways with matching there is no library, just software that makes connections between active participants.
‘Dumb’ matching systems just file providers and offerings in a passive library according to various attributes - they don’t actively seek out likely requesters.
An Artificially Intelligent matching platform understands not just what to do with user queries, but also what to do with the provider offerings.
This is a big advance for Machine Learning-enabled marketplaces over online ad listings or auction sites. In those, the seller uploads a lot of info that is categorised into a database and the user’s query is used to access that central resource. With Machine Learning matching platforms, the provider’s offering is ‘understood’ by the system rather than being filed in a database and then being accessed as a result of predetermined user filters.
One of the big problems with putting the right requester and provider together is that there are usually many possible answers and paths a user can take. This issue gets worse the more the requester’s need is “obtuse”.
Netflix is such a fascinating recommendation engine because many of its requesters have a highly obtuse need: they are looking for the right ‘something’ to distract them from boredom. The target the recommendation engine must hit is vague and arbitrary. The offerings it can present to make a connection number in the thousands.
And yet, through iterative improvement, Netflix is getting quite good at delivering a “signal” in this noise. Each time one the recommendations generated by its matching engine is selected by a user, it is a success that the system learns from.
There are many Machine Learning algorithms your matching platform could employ to analyse your data. In general, there are four main kinds of data they will use:
You’re going to be surprised and it is up to you to respond. At Netflix, for example, the company is investing US$5billion into creating new shows - some of which seem really left field.
You can bet that advanced algorithmic analyses of what performs well across the viewing behaviour of its 75 million users is the controlling factor into what programs are made and then picked up for a second season. The company is responding to what its relentless data acquisition finds.
In rule-based search systems, however, the parameters by which assets are categorised and by which results are returned are intentionally created by the software developers running the system.
The behaviour of the system is much less emergent and offers up far fewer surprises - in a way it can only show you what you are looking for, not what is actually happening.
The short list above is just a few things that you have to know about Matching Learning as it applies to matching platforms and marketplaces. It’s a field that is developing very fast and, as you can see, holds great promise for those looking to find new ways and new sectors in which people with needs and people with offerings can find each other.