This article kicks off the first of a series, where we delve into the world of Machine Learning and Visual Recognition. This article is an easy read for anyone with a basic understanding of PIM, while our upcoming ones will dig deeper into and share the best practices and code examples for inRiver PIM paired with IBM Watson and Microsoft Computer Vision.
During my 7 years and counting at Avensia, I’ve had the opportunity to deeply understand the wide range of needs of my customers; some of these customers are small scale entrepreneurs who know their inventory in and out, while some are large scale retail chains with a seasonal turnover in the tens of thousands.
While our customers might have very little in common in terms of scale, turnover, and markets, there’s one common denominator that they and their grandmothers stand to benefit from: good old Artificial Intelligence.
But don’t take my word on the serious potential of A.I. Get it straight from the horse’s mouth:
“Artificial intelligence is the future, not only for Russian, but for all of humankind. It comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.” - Vladimir Putin, currently ruler of only Russia
PIM is all about supplying the one source of truth, with regards to accurate product information. In essence, PIM supplies the ‘most precisest’ information of one single item of merchandise across a multitude of contrasting channels in a timely and SEO optimized manner, from the moment the product was entered into the database, to the moment after the customer clicks confirm purchase and the product has left the warehouse, to the moment the customer signs for the package, and to the moment when the product is re-entered into the database after the customer returned the product.
Entering this information has typically been performed manually. Having product information provided to us in an ungodly format from suppliers was our soul-churning reality, and keeping track of the ‘most current-est’ excel sheet involved meticulous organization, and luck, hope and a prayer.
We of course tried to enriched our product data to the best of our abilities, knowing that every little bit matters. Ensuring that product information is firstly entered into databases in a timely manner, and its accuracy constantly maintained, is a very expensive process in terms of time, finances, and often, my mental well being and general happiness. So, I’ve frequently pondered over and wondered about whether it’s just a far fetched dream to one day utilize A.I. to enter just enough information to let the product join at least a couple of our channels before any human has started writing a proper description.
But, today I can say that it’s not just foolish dreams, and working with Visual Recognition will be what transforms them into reality. Where we, early on in the enrichment chain apply Artificial Intelligence so smart, that we feel confident enough to let it do its work and see our products added to the customer’s cart before we can say Skynet.
Early bird gets the worm
The enrichment process varies from company to company but for my clients we often start out with basic ERP data, and if we are lucky the relation between a product and its SKUs. For many the next set of information are images, either from their supplier or from their own photography process. The best part is that both matter, even if the supplier images are to be swapped out at a later stage.
What we do, is that as soon as we have the product set up in our PIM and it has that one single image - we let A.I go to work. This will yield the greatest benefit, by either adding information aiding the editors in PIM or even enriching it enough making it ready for any or all channels.
Multiple images will of course help immensely but they are not necessary to start the process. They are evaluated one after the other, never in multiples. We have to apply basic math such as mean to determine the result. But for this to work, we need to start with what we already know about our current inventory.
For Visual Recognition to work we need to supply images with related data. This data should be in the form of keys; data we can easily match in our PIM system to information about the product. And good data to start out with are our predefined key value pairs available on the product in most PIM systems.
Let’s start out with an example - Color. For my last client we implemented automatic calculation of RGB color codes, taking into account background color. This worked well for swatches, but it failed miserably on images where the actual color was a small part of the image, or where the product was multi colored. In other words, this data was useful for displaying colors on the product page, but was ineffective for searching or filtering products.
What if we could teach our A.I that the color we are looking for is not the white background, not the black bottle cap, nor the red label - but rather, that the color we are looking for is on or in the contents of the product? And what if we wanted the result to be in a set predefined colors we organize our products by, so as to not have two hundred filtering options for color on our site, would that be possible? Of course, it’s possible!
By supplying our A.I with images from our current product catalogue where color is already defined, we teach it which parts of the image are of interest to our result and how to properly determine the desired result. And it works quite effectively. Actually, this works so effectively that I’ve switched from staying awake at night by replaying my daily the nightmare of organizing product information to worrying about being out of a job.