Humankind has been fascinated by objects and organisms that exist in an eternal loop, like a torus or the mythological Ouroboros – self-sustaining systems in the purest sense.
If you’ll allow the conceptual leap, this pattern has similarly emerged in the way that digital products are being developed today – a symbiosis of consumption and the harvesting of user data.
A neat example of this is Spotify’s Discover Weekly feature, where 30 new tracks the user hasn’t listened to before are served up every Monday. How does Spotify work out what tracks will pique every individual’s interest? Well, based on the music a user listens to, Spotify assigns them a taste profile. Spotify scans millions of playlists, building connections between tracks, and matches those to individual taste profiles. The end result is an experience for the individual of constant discovery and, more often than not, music satisfaction.
How the Discover Weekly machine works (inspired by Nikhil Sonnad, Quartz, & Fabien Girardin BBVA)
When it works, it’s like magic
Unpicking it all a bit more, there’s the thin slice of data that every individual user is creating and there’s the fat agglomeration of all the data created by the entire community. When these are combined and fed back to the user, it enables another virtuous loop.
To enable this functionality, we come across two types of machine learning problem, namely that of clustering (to determine which songs go together), and recommendation (to create the list of 30 tracks).
Whilst we might be more familiar with the applications of these technologies in the media and ecommerce sectors, we will increasingly see their application in the world of consumer financial services. Everything from analysis of the evolution of savings patterns to determine which investment products are more appropriate to recommend, to the more extreme of providing recommendations of where to eat whilst on holiday based on spending behaviour blended with third party ratings.
Machine learning techniques can be incorporated into digital products and services in a plethora of ways, but just like any technology, it is not a panacea – and it must ultimately fit with the product goal and business goals.
As part of an analysis process to decide how machine learning will fit in, it’s useful to ask yourself a few questions that can help frame your thinking:
- What are we trying to minimise or maximise for with this product or feature?
- eg. for a money transfer feature, auto-suggesting friends to send money to not only helps reduce the time taken for the user, but also the cognitive load.
- What data is actually available and can we first do this without machine learning?
- eg. do we have access to the user’s location data, or their spending data?
- What is the velocity of the data collection and presentation to the users?
- eg. Spotify’s Discover Weekly only sends out recommendations once a week, whereas an app like Yelp that can notify you with good restaurants nearby that match your tastes, must be able to adapt in real time.
- How contained and clear is the experience?
- eg. is it limited to a particular piece of singular functionality and can it show its “seams,” such as Skyscanner predicting whether the price of a flight might go up or down in the next week, providing a confidence score with it.
- What kind of feedback can end users provide, in a natural way, to improve the performance of the machine learning model?
- eg. the facial recognition feature in iOS photos app allows the user to easily correct its grouping of photos with an individual in it.
For the financial services sector, the potential to create digital products and services that are self-reinforcing in terms of the experience they deliver, by the data generated by users, is huge.
But as with everything, you need to start simple, only gradually building on top of this as the fruits of the initial project becomes clearer.