Recently I had a few meetings in which it became increasingly clear to me that even long-standing customers do not know exactly how we generate the customer profiles for our personalization. I hope I have figured out a suitable explanation now.
First of all it has to be understood that each of us has a personal taste. This changes, but every change is a flowing process that can be analyzed and recognized with sufficient data.
Now to the data: Each customer receives an anonymous customer ID. This ID is assigned to purchased items, but also the products that were viewed and clicked (as a time series). In this way, we learn what the customer looked at, in which order and for what he then decided.
If we combine this data with the shop’s entire and relevant assortment and the appropriate external trends, we get a pretty good picture of the customer’s preferences. For all products, we then export the visual data (we call these features, which are vectors from floating point values), combine them using a sophisticated dimension reduction algorithm, and then get a taste fingerprint. This “fingerprint” is then itself a vector of floating point values.
We could use this fingerprint to compare the preferences of customers with each other or with offered products. Every product has a fingerprint that we can use to see how well it fits a particular customer.
We combine three very interesting disciplines: image recognition, dimensional reduction and search. Each of these disciplines is exciting in itself, but together they not only provide an extremely efficient way of personalization, but also a huge field for further research.