Today I did a talk at the San Francisco Agile Marketing meet up. It was geared to a non-technical audience, and I hoped I help guide everyone to see how they can use machine learning with their online store. Machine Learning's not just for the big guys anymore.
David Jones speaking at the SF Agile Marketing meetup.
Here's the video from my talk.
And here are the slides.
If you've got an online store and are interested in using machine learning to improve the customer experience and increase sales, please do get in touch to discuss your project.
Machine Learning and Ecommerce: Talk Transcript
Whether we realize it or not, we're all familiar with machine learning. Consider Netflix. Who takes your preferences and your watching history and recommends movies to you. And Gmail: Billions of emails analyzed to keep spam out of your inbox. And Amazon: harnessing the power of your purchasing and browsing history to recommend products to you.
Let’s take a look at Amazon. This is Amazon's home page this morning for me. Aside from learning that I'm interested in running and coffee, what we can see if we blur-out everything that hasn't come through machine learning, you realize that about 50% of the page has come through some sort of recommendation engine.
This is a big company taking the fantastic benefits of machine learning and generating a lot of revenue from it. The thing is that machine learning is not just for the big guys anymore.
Never before has machine learning been available to all of us. So today I want to talk about machine learning specifically for e-commerce. These are some of the things you can do with e-commerce.
There’s email marketing. Why are we still sending the same email content to everyone when we send an email campaign? With machine learning we can individualize those emails and send relevant content to each person rather than sending everyone the same message.
And cross-selling. The ability to take a product someone is looking at and recommend similar or complimentary products.
Personalized recommendations. Based on everything I know about you and what you've looked at it, maybe there are products I think you'd be really interested haven't yet seen.
Abandoned cart emails. If I added something to my cart and haven’t completed the order and I close my browser. Maybe I got distracted at the last moment or maybe I decided the price was too high. With machine learning we can wait an ideal period of time and send a personalized discount to that person later on, trying to convert that person to an actual sale.
Finally there’s lead scoring. How do we know which customers we should be focusing on? With lead scoring we’re able to figure out which customers may have not purchased just yet but are very close to buying. In this way we can focus our support and business development people on those prospects.
Let's assume you're interested in integrating e-commerce and machine learning into your store. There are three steps to do that. First we need to collect some data. Second, we need to train a model. And third, we can make predictions.
In the first step we're talking about collecting data on purchases, products views, likes and ratings, and wish lists. It's really any data that’s connecting customers to products.
Second, we need to train the model. For this step I like to use a system called Prediction.IO. It’s an open-source machine learning server that takes the data you collected in your store and provides an API for machine learning. Your developers can take that and integrate it with your site. It also comes with several ecommerce templates so many of the use cases I’ve discussed you can get up and running right out of the box. Finally it’s production-ready. What I mean by that is it’s easy to set-up, easy to maintain, and it stays reliable.
Finally we're at the fun part: Making the predictions. And for this I want talk about a real-world example from one of our clients: United Cellars. They sell wine online to customers in New Zealand and Australia. We thought we could use machine learning to increase their revenue.
The data we had to work with: 16,000 product views, 60,000 orders and 3,000 product ratings. This is the data connecting products to users. We’re talking about 79,000 rows of data total. By Bay Area standards this is pretty small.
Another consideration is that with wine, everyone has a diverse taste range. United Cellars has over 10,000 wines so machine learning seems like a really good fit. We have a lot of products and each one of us is very individualized in our tastes. So it's about creating something that's going to bring those things together.
We decided on implementing two things: Personalized recommendations and similar products. With personalized recommendations this is recommending, out of everything in the entire store, which products we think are most relevant to that customer. So it's very much related around the customer’s context.
The second thing is around similar products. This is in the context of the product itself; If I'm looking at a product, showing me products that are related to that product. It doesn't really care who I am, but rather what I'm looking at.
We decided to run an A/B test to see how effective this was on the site. This is the home page and you can see across the left hand panel we've got “recommended for you,” that's the personalized recommendations. Along the bottom we've got “more red wine to consider” which is the similar products.
When we put this live we realized very quickly that we were starting to recommend products that were out of stock. You have to a very considerate about anything that would lessen the relevance of your recommendations so we added a filter to make sure we don't show any out of stock products.
The second thing is that you have to be careful about understanding the context of the business you’re operating in. With wine, people tend to prefer red or white wine, so if I'm shopping for red wine show me more red wines, don't show me a dessert wine or something else. There's a much higher chance that I’m interested in red wine.
The results: We ran this for a little while and we got 45% longer average session, 22% increase in conversion rate and 37% increase in average order size. This generated a 71% increase in revenue.
This is huge. It’s because we had an increase in conversion rate and an increase in average order size. Those two things can be very very powerful for revenue.
Machine learning generates revenue but it also improves the customer experience. If I on a site and I'm searching around it can be hard work finding products that might be interesting. If machine learning can help us bring customers closer to those products more quickly and delight them with things they wouldn't have looked at-- that's definitely a better customer experience.
Three things I learned. The human touch is still important. You must understand the context of the business and figure out which use cases with machine learning may make sense. If you have a store that has just six products for example there's no point having personalized recommendations or similar products because, as a user, I can already see all those products. So perhaps picking one of the other use cases makes sense. But it takes a human to understand these sorts of things and how the shopper shops and figure out exactly what to show them.
The second thing is that Prediction.IO is awesome! It makes it so easy for all of us to start using machine learning and setting up one of these servers that can take data in from your store and start serving predictions.
Finally, small data can be enough. It’s something we don't hear about that much in the Bay Area. In my example we had 79,000 rows. It's really not the point about number of rows but rather the number of connections between the products and the customers. So if I had half the number of products, half the amount of data would be okay provided I still had those connections.
You can go to https://resolve.digital if you want to see the full case study for United Cellars. Thank you.
Join The Conversation
More On The Blog
Are Spree sites meeting Google’s page speed standards?
A study of 680 sites built on the Spree OS platform reveals nearly all are mobile friendly, but only a handful achieved a mobile speed score high enough to be considered by Google as performing well.
Barry Harrison — Sep 12, 2016
Machine Learning and eCommerce Talk at SolidusConf
I recently spoke at SolidusConf about Machine Learning and eCommerce
David Jones — Jun 14, 2016
Top 7 Product Recommendation Options
Top 7 Product Recommendation Options Ranked by Revenue Generation
Barry Harrison — May 31, 2016