Today, most recommender systems are based on a class of methods commonly called ‘collaborative filtering’ - which means that they generate recommendations based on a user's past behavior. This method is successfully used by Amazon and Netflix (see the https://en.wikipedia.org/wiki/Netflix_Prize). They are also very unsuccessfully used by smaller companies that lack the critical mass of historical behavioral data required to use those models effectively. This generally results in the cold start problem (https://en.wikipedia.org/wiki/Cold_start_(recommender_system...) and a worse customer experience. We solve this by not focusing on understanding the customer but instead focus on understanding the product.
The way we do this is with machine learning techniques that create vector representations of products based on the products’ images and descriptions, and recommend matching using these vector representations. More specifically, we have found a way to scrape the web and then train massive neural networks on e-commerce products. This makes it possible to leverage large amounts of product metadata to make truly impressive recommendations for any e-commerce store.
One analogy we like is that just as almost no single company has enough sales or behavioral data to consistently predict, for instance, credit card frauds on their own, almost no e-commerce company has enough data to generate good recommendations based only on their own information. Stripe can make excellent fraud detection models by pooling transactions from many smaller companies, and we can do the same thing for personalizing e-commerce stores by pooling product metadata.
Through A/B-tests we have proved that we can increase top-line revenue with 4-6% for almost any e-commerce store. To prove our value we offer the tests and setup 100% for free. We make money by taking a cut of the revenue uplift we generate in the A/B-tests. We have also found that the sales and decision cycle gets much shorter by being independent of customer's user data. You can see us live at Staples Nordics and kitchentime.com, among others.
Oliver and I have several years of experience applying recommender systems within e-commerce and education respectively and felt uneasy about a winner-takes-it-all development where the largest companies could use their data supremacy to out-personalize any smaller company. Our goal is to build a company that can offer the best personalization to any e-commerce store, not just the ones with enough data.
Do you think our approach seems interesting, crazy, lazy or somewhere in the middle? We’d love any feedback - please feel free to shoot us comments below or DM, we’ll be here to answer your thoughts and gather feedback!