Raphael Geronimi, Machine Learning Expert
Mike Cleary: Today we are sitting down with Raphael Geronimi to discuss the topic of machine learning. Raphael is a specialist in the design and implementation of machine learning algorithms. He has delivered machine learning solutions for web advertisers for real time bidding applications with over 200,000 online predictions per second, financial organizations providing credit scoring and fraud detection, social networks providing recommendations across millions of users, and e-commerce merchants. He has also served as a big data and machine learning advisor to startups including Fusion. Thank you for sitting down with us today to answer a few questions.
Raphael Geronimi: My pleasure.
Mike: Let’s start with the language we see out in the marketplace. The terms ‘machine learning’ and ‘artificial intelligence’ are being widely used by companies to describe how they are using data to better predict behavior. How would you specifically define ‘machine learning’ vs. ‘artificial intelligence’?
Raphael: That’s a good question. There are several conflicting definitions of artificial intelligence and machine learning. The one I think is the most relevant is that machine learning is a specific subclass of artificial intelligence techniques. In the past, the most favored approach was very different and what I would call ‘created logic’. Humans would kind of ‘hard-code’ logic rules and through brute force exploration, use computers to find the most efficient combinations of these rules to reach an objective.
The typical case 20 years ago was chess games where you would play against a computer. That was ‘created logic’. Someone coded all the rules and then the CPU would go through all of the combinations and find the best combination. Machine learning is very different because it is not ‘created logic’, it is ‘created data’. Some humans have decided how to feed the system with the right data and then the system finds the rules by itself. This is much scarier but is more impactful for businesses.
Mike: Interesting. We see confusion out there in the marketplace around machine learning algorithms versus a hard-coded yet sophisticated personalization rules engine. Can you elaborate more on the difference between machine learning and a sophisticated rules engine?
Raphael: Yes, a good example follows what I just said about the chess games. In the world we are in now, compared to 20 or 30 years ago, there has been a data explosion. There are terabytes of data across all applications we use on a daily basis – and that makes ‘created logic’ and hard-coding rules very hard. It doesn’t scale. You need humans to look at everything and we know that’s not feasible. With machine learning, we are able to scale much more efficiently to find the needle in the haystack with 99% automation and only one person or human interaction to just make sure that it works. It’s the new era we are in: the New Data Era. It’s about scalability.
Mike: A lot of people and companies are using the term “machine learning” or “artificial intelligence” to describe what they’re doing. In your experience, how prevalent is the use of true machine learning algorithms at scale in production environments and what kind of companies and what industries are truly using that capability?
Raphael: That’s a good question. We see these buzzwords being used everywhere on a daily basis. The truth working with many companies is that in many places it’s very primitive. When you look at exactly what many businesses are doing, it’s not so different from what happened 20 years ago. You often see hard-coded rules. Sometimes you see some basic statistics. It is pretty rare that today you see machine learning being really deployed. It doesn’t mean that it doesn’t work — it’s pretty much the opposite – it’s that companies do not realize how well it works! And because they don’t see it working internally, they don’t realize that they are missing a big opportunity to use it and therefore have much more impactful systems.
Mike: So here at Fusion, we work to optimize the sale of ancillary or add on products currently with a heavy focus on the travel industry. Can you talk about why this may be a good fit for the application of machine learning algorithms?
Raphael: So, I think there are a few patterns which make machine learning particularly efficient compared to hard-created logic. One is the data volume. This is something that we see in the travel industry: the volume of users going through websites comparing prices, selecting a flight, and then going through all these details. Add to that, data like the travel history of the person and all the details of the flight and the trip. It’s a big opportunity! Leveraging all of this data to be able to find the right price for the user and to automate the right recommendation in terms of ancillary product purchase is right in front of us and it’s important to leverage it.
The second case, which makes it particularly appropriate for machine learning, is that the data is not motionless. We are talking about what we call “continuous data” and things which you can put on a scale like numerical values such as the number of miles, the distance of the flight, or the number of miles accumulated on the account of the user. And then you have “categorical data” like describing which destination airport. All these things are intelligence so it’s very hard for a human to make sense of all of that and say “well, in that specific case, we can put this specific product.” You cannot do that with humans; even people with a lot of experience in this industry! That’s where machine learning is particularly helpful. Lots of heterogeneous data is a big opportunity.
Mike: Tell us a little bit more about the types of machine learning algorithms that are available, specifically the bandit class of algorithms. I understand these can be used to optimize recommendations based on data that’s collected during a transaction that can be used to present custom offers to the customer. Walk us through what these algorithms are actually doing with the data that’s coming in as well as matching that up to the different types of features and alternatives of those features that then get presented to a customer.
Raphael: Sure. So when a new customer arrives, you do have some data points and history about the customer, their profile, the flight they are looking at, all these details — and you have alternative things or actions you could perform on this user. You have some flexibility: You could propose to them some products or specific prices on those products. The question becomes: I have all this information about the user (which we call the context of the user) and I have all these alternatives, what should I do?
And to answer that question, there are two things you could do. You could do exploration or exploitation. Exploration is how you try something new on this user. I don’t know if it will work, but this will allow me to see how they will react. For example, I have this new pricing point that is now possible and I want to see how the users react to this new price: will they react well or not?
Another thing you could do is exploitation which is basically not trying to see how the user reacts (you already know that, you have already done some experiments in the past) but exploiting an opportunity because you know this will work. You know that the specific price point is a sweet spot for these users and then you want to propose it again and again.
And these two things are kind of contradictory because exploration means you accept the possibility that you might lose some opportunities with user. Exploitation means you accept to lose your opportunity to discover new things about users like new reactions to price points or specific offers you could make. Finding the optimal balance between both things has been a challenge for many companies and is specifically the question that is answered by bandit algorithms. Bandit algorithms are a sub-part of machine learning, of reinforcement learning. They focus exactly on navigating the optimal trajectory for you. They will balance, in an optimal way, exploration with exploitation to guarantee you that, over a given time span, you will get the optimal “performance” you can get from your customers. (“Performance” being measured with business measures, not statistical measures.)
Mike: Finally, all this talk of machines, machine learning, artificial intelligence, may lead some to believe that the machine can simply run itself and humans are no longer required either from management of the machine learning algorithms themselves or from a marketing perspective. What are your thoughts on the role of experienced marketers working with machine learning experts and how the best companies marry these two types of expertise in both of these areas to achieve the best results?
Raphael: I think it’s a very important question because I see many false promises made on machine learning that it will automate all the parts of your business and I think that’s completely wrong. To the contrary, machine learning is a huge opportunity for companies to focus on what they are best at.
I think as the techniques stand today, humans have an unbeatable advantage over the machine – which is the creative component. They know the industries and verticals very well. They have some very deep ideas about the profile of customers coming to a website and what they are looking for. So they have a natural advantage over machines which means that they are able, based on their human experience, to have opinions about what might work but they don’t know if it will work. They don’t know how far to explore versus exploit.
How much of a budget you can put to this becomes a difficult task for humans, but they have the right ideas. They have the creative part. And I think to the opposite, machines and machine learning and bandit algorithms are very good with the objective part, which is: given all these things we want to try, which makes sense to try on our customers and which are the ones that work in practice? Machines are excellent at finding the optimal trajectory and based on facts (and only facts) dropping one alternative after the next until you know exactly what works. And the optimal coupling for me is when you have marketing teams focused on just generating new ideas, leveraging their experience and their understanding of their customer’s needs. You couple that with bandit algorithms which manage the objective part. They do experiments mechanically on a very large scale automatically and they kill one alternative after the next to pinpoint exactly what works.
Mike: So rather than replacing the human element, the implementation of this type of technology can actually result in the need for more focus and more people on the marketing side.
Raphael: I think machine learning allows us to really separate the functions. There is a creative function and there is an objective function. And I think it’s a huge opportunity for marketing teams to really focus on things they are good at. This has been a difficulty for many businesses in the past, you have some theory of your customer and your business — How do you test your theory on a large scale and make sure that this is a reality? Many businesses don’t have the ability to wait three years to see that this worked or didn’t work. They really need to be informed pretty quickly that things behave as expected or not. And this is an opportunity to do that. And to adjust course, discovering how the market reacts.
Mike: Thank you Raphael for joining us today.
Raphael: Yes, thank you for the opportunity!
Mike has over 15 years of experience in marketing and technology spanning the areas of product management, digital marketing, marketing analytics, market research, and brand strategy. He holds a Bachelor of Science degree in Commerce from the McIntire School of Commerce at the University of Virginia.
Latest posts by Michael Cleary (see all)
- SHOP TALKS #3: Interview: Raphael Geronimi, Machine Learning Expert – 20/March/2019
- The Future of Customer Experience Management – 6/March/2017