April 6, 2020
Ecommerce’s machine learning evolution

Ecommerce’s machine learning evolution

Hi! I’m Hans-Kristian. I’m the founder of
Clerk.io and today I’m gonna talk about machine learning in e-commerce. In the very beginning, you can say that the basics of any kind of machine learning technology to recommend products in e-commerce that started, it’s easiest to say, that just started with counting. You count up and just say “ok I have this product, what would go well with it? well, that’s what I have sold most with it” A basic thing like counting is the easiest way to understand where all the machine learning algorithms came from. The problem with counting is: “ok, all your accessories that are sold with everything, they start popping up everywhere. Because they are sold a lot, and with a lot of products, even though they’re not
relevant.” And to solve this problem Collaborative Filtering was introduced as the first real effort to make something good in recommendation technology. Collaborative Filtering basically says “ok, I have a customer who have bought these products. What other customers have bought the same products? and what have they bought beside it that might be interesting for the next customer? It’s a good idea and it’s simple to do but it requires a lot of data. It requires a lot of purchases by a lot of
customers in order to get a good fit. So, the problem is that you will see it might work with the top buying customers and the most purchased products might have good recommendations and the rest might be so so and you have to do a lot
of manual tweaking to actually get it working. The next step on top of this which is trending right now is in Neural Networks. With Neural Networks
customers don’t have to actually purchased the same products you can be a bit more fine-grained; meaning that, the Neural Network should
get better results without as much tweaking but it requires still a lot and a lot of data. so you can still go in and do a better job than with Collaborative Filtering but Neural Networks gives you a better results. Then we come to what
we do because a big problem with both Collaborative Filtering and Neural Networks is that you need a lot
of data in order to get good results and that’s basically the problem we solve. We solve the Cold-Start problem. so we can deliver the same kind of results as Neural Networks but with almost no data and the way we did this was going in to figure out: “ok, instead of looking at individual customers and comparing those; instead of comparing the individual products in the store, we have a base layer of technology that understands the store as a human would do” So, if you have a a toy store we’ll go in and figure out which brands go well together in general and then the basic things that you would know as a human being if you have a clothing store. we understand that colors are important here where in the toy store colors might not be important and
relevant. We have this basic understanding that you have as a human
and we deduct that from the data and we use that on top of all the orders, on top of all the products to actually fill out all the gaps and have a near human level of understanding of what people want to buy. I hope you learned something about Recommender systems in e-commerce, it’s a big topic, it’s difficult, but it’s fortunately really easy to try out so you can just click
down here for a free trial of Clerk and I hope to see you soon

Leave a Reply

Your email address will not be published. Required fields are marked *