April 10, 2017 / by Elle Panes
We open the blog to Michal Szczesny, COO at Artfinder. Artfinder is the art marketplace which connects people to artists and art. Michal will be speaking at our event in London on April 11th. The last few tickets remaining are available here.
We’re an art company. Or at least, we used to be. That’s not strictly true – Artfinder has always been a tech startup first and foremost, but we’re a tech company that is providing a solution to a very human problem – connecting independent artists directly with customers around the world who would love to buy original art, but either didn’t have access to artists or didn’t know they could afford originals (which are very affordable when you take out the middleman – i.e – the gallery).
We first started investigating AI when we realised that we reached a critical volume of work on the site (we’re at 350,000 pieces now, that’s more than MoMA). A vast majority of those are unique, one-off pieces that cannot be recommended multiple times after they sold. That amount of art is brilliant and exhilarating, but we can also see that for users it could be overwhelming. We’ve always had search filters on the site, plus text search, but to search for a ‘large blue landscape painting for £100 – £200’ and to find 50,000 results isn’t an especially encouraging experience. There seems to be this mindset that once you’ve put in search criteria you have to look at *every single result* to make sure you find the best one – and most people would rather give up than look through 50+ of pages of search results.
So, this left us with the ‘needle in a haystack’ problem – the need to match customers up as quickly as possible with a manageable selection of artworks that they love. For us, this problem is compounded by the ‘I know it when I see it’ mindset, which is that shopping for art is not like shopping for shoes or books, where you usually know what you want or can describe it with words. Art shoppers tend to start from either a complete lack of knowledge about what they want or a with a visual idea which is very difficult to put into words.
We knew we had to solve the ‘needle in a haystack’ problem in a creative way. And art taste is such a subjective, personal topic – we want our users to feel that Artfinder knows and understands your taste, without ever being judgmental or intimidating, like an art gallery might be.
We began our AI journey with personalised recommendations driven by machine learning and graph database software, Neo4j. From 10,000 artists, 350,000 artworks and 600,000 users we have a huge amount of data to feed into the graph, plus several different levels of relationship between products, artists and users. Users and artists can view or ‘love’ an artwork, ‘follow’ an artist or user, or of course buy an artwork. Those relationships can then be weighted (purchase is stronger than ‘love’ for instance) and we can calculate product recommendations for you based on what similar users have liked.
Those personalised recommendations are great, but obviously you need to have got as far as looking at an artwork on the site before they become useful. In the meantime we had also built an onsite feature called ‘more like this’ – which uses open source visual similarity detection software called LIRE to ‘match’ any artwork on the site to up to 400 others. The same software is used by the police for facial recognition from CCTV cameras. It looks at visual structures on a deeper level than similarity that could be described with text like colour or subject matter.
But we still needed a way to engage users who weren’t even on the site, and who didn’t know what they were looking for beyond a visual stimulus.
Emma fulfilled all of those criteria. Users can tweet any image at her (a photo, an artwork, a selfie) and she will reply with Artfinder artworks that are inspired by your image. She’s a really easy, good fun way to dive into our catalogue without having to even think about what you’re looking for. She’s also, of course, a little bit cheeky – there seems to be this conception that buying a piece of art is a very serious and considered thing, something you might need an advisor for – and what we’re trying to do is shake people out of that.
Buying art can be all of that, if you want it to be, but it can also be as easy as buying a pair of shoes, or a cushion or a pot plant. Art will always be valuable, and will always be something to love and care for, but that doesn’t mean it has to be expensive. Emma is helping us communicate that to our users.
As we go further and further on our machine learning journey, Emma will recommend art you’ll love with (hopefully not too) scary precision, drastically reducing the ‘needle in a haystack’ problem.