by Joseph Kibe, Head of Data Science at Proton.ai
Here at Proton, we’re often asked what makes our recommendations so great. It’s a good question: the recommendations that we provide form the core of our product. We can make our systems as easy to use and to integrate as possible, but without good suggestions for our users, we’re sunk.
The key lies in a few choices that we’ve made. We spend a lot of time talking to our users and industry leaders, so we understand what kinds of recommendations are useful, and what kinds of adjustments need to be made to suit the business environment we work in. We also make our systems flexible and performant: they’re fast and we can improve them more quickly. And finally, we use cutting-edge machine learning algorithms, so our customers aren’t getting warmed-over technology from the 1990s.
It’s that third area I want to consider today. At Proton, we use neural network models, which offer huge advantages over previous-generation machine learning techniques. They take their inspiration from the human brain and its network of neurons, and are largely responsible for advances in artificial intelligence products in the last decade.
But what makes them so great and so much better than previous models? Unlike previous-generation models, neural networks are capable of learning non-linear patterns. That may sound complicated — and in the details it can be — but it’s actually pretty easy to understand.
Let’s consider a simple example. Suppose you’re running a bookstore and want to recommend the next book customers should buy when they come visit.
With older techniques, we can only really consider a particular factor in a simple, linear way. We could draw simple boundaries. People under 12 get recommendations for kids books, people over 12 get recommendations for best-selling novels. People who buy mystery novels get more recommendations for police procedurals.An example of a rules-based recommendations. This model predicts that a customer’s likelihood to buy a children’s book is linearly correlated to their age.
This might work some of the time, but it doesn’t take an expert in books to realize that it’s often going to be wrong. What if a parent comes to the store? Do we suggest a book for him, or for his children? What if someone likes political mysteries? Do we suggest a mystery novel or a book about politics?
With a neural network, we can account for strange exceptions that rigid, rules-based recommendations can’t. We can accommodate the fact that, say, starting at age seven, some children want to start reading serious novels. Our confidence goes up in a strange pattern up to the point they’re when they might be buying children’s books as gifts, then ticking up again, even more, when they’re over 60 as they might become grandparents. It’s impossible to encode this complexity with older techniques, but with neural networks, we can.An example neural net recommendation model. This schema is based on real-world behavior and finds that even if customer behavior is predictable, it is not linearly predictable.
Older models may be better than choosing at random, but they can’t capture the richness and complexity of the real world. On average, children are a lot more likely to buy children’s novels than adults. But if the bookstore applied simple rules, they might oversell children’s books to teens, and miss giving parents and grandparents accurate recommendations for their children and grandchildren.
Despite appearances, not all recommendations systems are the same. While it’s likely that adopting any kind of predictive technology will have a business impact, how that system works under the hood makes a big difference to the size of that impact. A good bookstore has all the most popular titles in stock. A great bookstore has what’s right for you, and helps you find books you didn’t know you needed. With good underlying technology, we can bring that same experience everywhere.