Corey Haines and Daniel Smith joined forces in this podcast to talk about machine learning and how it’s useful for marketers.
Daniel Smith is the Technical Architect of Cordial, bringing his wit and years of experience to lead the machine learning initiatives that drive many of Cordial’s unique features. Daniel hates the word “machine learning,” but doesn’t shy away from why it’s so important for marketers. How is machine learning driving some of the most impactful technology in marketing? If you’d like to listen to the audio, you can find links at the top and bottom of this post.
The following is the podcast transcription in a question and answer format.
Corey: Hey guys this is Corey from Cordial and you’re listening to the Be Cordial podcast where we share stories and insights on all things marketing and I have with me Daniel Smith the technical architect of Cordial. Today we’re really talking about the reality of using machine learning to improve email. But first Daniel, why don’t you just give us a brief review of your background, your experience, and some cool things you’ve done?
Daniel: Hey guys so I have been doing this for a long time. At least for me a long time more than 15 years and 15 years of data management data workflows, dealing with all kinds of data problems that people have had in the marketing world. And so I kind of lived through the rise of big data as a concept and then the absence of big data as a concern as those things played out and were not a challenge as much anymore. And kind of naturally all of those things kind of led to what I’m doing today with Cordial.
Corey: So yeah that’s great. Well let’s get into machine learning. So it’s a buzzword, we all know it. No one knows what it means, is it artificial intelligence, is it machine learning. Why don’t we just define what is machine learning.
Daniel: Yeah and maybe I would define it as what is machine learning to Cordial a little bit too because I think everybody takes their own perspective on it. I mean machine learning and I frankly hate the term I think it’s overused.
We use it. I think it is a good indicator of what’s going on or the covers but at its core machine learning is the use of algorithms or mathematical equations to gather insights from data and allow machines to learn and make decisions about the data and potentially provide insights that weren’t visible to individuals looking at the data because there’s so much of it that kind of plays into the big data that I mentioned in my background. When there’s that much data you kind of have to let the machines do the work and so the machine learning fundamentally is just math applied to a lot of data.
Corey: Yeah. And so as far as I mean what a computer can do what a machine can learn… Are we talking like a child’s intelligence or are we talking human intelligence, talking superhuman? I mean what’s the level of intelligence?
Daniel: So there is deep learning which is the next layer of machine learning we don’t really get into that but every complex system has kind of a set of inputs and outputs. And I mean you think about any system in the world whether it’s technical or not. It has an input that does something and there’s an output and machine learning it today, at least, and I’m going to disregard deep learning for a moment is really only as good as the data you give it.
And so there’s a lot of work you have to put into understanding what is the right data to give it, what is the right kind of algorithm to apply to that data. And that is the output what you’re looking for and there’s a lot of testing of that to see if if you get what you expect.
But all that to say it’s intelligent and I think one of the reasons we think of it as intelligent is because maybe we’re surprised by the insights that we can kind of garner from that data that we wouldn’t have otherwise done. I think as humans we kind of have this perspective that we’re kind of the you know Master Intelligence race, whatever.
And and we use that maybe to our advantage in a lot of situations where we’re able to make decisions and and glean insights from data and see patterns really easily.
And we are surprised when computers can find patterns that we can’t find, but ultimately it comes back to data in data out. And so it’s only going to be as smart as the data you give it and then it can be even smarter on the way out because ugly things that you couldn’t otherwise have tell.
Corey: So why is machine learning a buzzword today. I mean it’s been around since the 50s and it’s come in and out. But why is it popular today.
Daniel: Yeah yeah. And it kind of it’s funny some of the math that we use even today is is from the 50s.
I think part of it is the accessibility of our ability to look at a lot of data and apply those kind of those algorithms to a lot of data. So Machine learning is probably a great way to classify all of those things.
There are supporting terms like artificial intelligence or data science that all surround the same field.
I think machine learning is maybe an easy way to describe it. But it has been used for a long time and I think a lot of those kind of enablement tools the public clouds like AWS or Google compute or others, have enabled people to use that data.
And so it’s become more accessible and not just to government agencies or contractors. And so it’s become kind of it’s been put in the limelight.
Corey: Yeah. You know it’s been around for a while. That’s no joke. And and you mentioned before that we use some of the same technology that was around since the 50s but how has it evolved in the last sixty seven years?
Daniel: I would slightly correct that and say it’s not the same technology necessarily but it’s the same math and.
And I think the evolution and I love just a little bit of the evolution is the accessibility of compute CPU. So the ability to buy that by the hour on public clouds you know primarily driven by Amazon and being able to look at that much data all at once, think about computers in the 50s… the math may have been there but the data storage and the accessibility of that to startups that can use that use those resources on a short term basis as opposed to having to put out a lot of capital to build up huge server farms.
I think a lot of that has enabled really smart people who maybe didn’t have the ability to get capital to go and test things that they wouldn’t have been able to test otherwise and so it’s created this rapid innovative cycle where people can use those resources and apply that math from 60–70 years ago to data today and get insights that they would have been able to otherwise.
Corey: So how is machine learning being used today, whether it’s Martech, whether it’s businesses, whether it’s product recommendations. I mean talk to us a little bit about how it’s being used and why it’s important.
Daniel: Yeah I mean I think fundamentally you can come back to a few different classes of products. There’s a couple that are a little bit more difficult.
There’s natural language processing which allows you to develop grammar and maybe develop a subject line in the case of an e-mail. There’s cohort analysis which is looking at similarities between groups of things whether that’s people that come to think of words as people but looking at similarities between products, people, other things, testing.
So moving away from a traditional probabilistic test and using other algorithms to get to that data faster so there’s kind of a testing component and then there’s recommenders which would be content based or purchase based and there’s a couple of core innovators in that space. But I would say that those are the main ways that they’re filtering into marketing.
Corey: Well let’s talk about e-mail specifically how. I mean let’s talk about Cordial as an easy example. How does Cordial use machine learning in the platform.
Daniel: The core machine learning offerings we have today are a cover testing, so testing of content, multi variant tests in a rapid way that minimize both cost and what we call regret. So a regret, is reducing sending the wrong person the wrong content. And so if we can get away from that and make sure that we’re sending the right person, the right content, more quickly.
Then, we can reduce that regret, reduce the cost of that, and ultimately the opportunity cost of getting some marketing in front of them. So there’s the testing side, there’s also recommendation side, so anything from recommending products based on likelihood to purchase on a lot of factors there.
We’re doing some really interesting stuff there.
And then on the cohort analysis side, figuring out naturally which people behave similarly even if as humans we don’t see it and one of the interesting things is something we toss around the office quite a bit and chuckle about is that we like to think that we’re all unique and we like to think that every human is is unique in their own way.
But the reality is and Netflix proved this with their movie recommendations is that there is somebody else in the world just like you with your same preferences your same purchase behaviors or same habits.
And so in that our goal is to go, “hey who are all those people who are similar and put them together in a group and say hey how do we identify these people.”
How do we make them look like each other or help identify what they look like each other and what’s working with that group. And then if we can segregate those groups into chunks of people maybe we can figure out a better way to market to each of them independently and have a better outcome.
Corey: Speaking of Netflix, who are some of the other major innovators in the space that are utilizing machine learning and using it to gain an advantage.
Daniel: The two that I think are the biggest and are probably worth mentioning right out of the gate are Netflix and Amazon. So Netflix has Netflix prize which was a goal to better their recommendations by 10 percent and it took them two years to get to 10 percent improvement and that was offering basically a million dollar bounty on some group of people or individual coming up with a better kind of recommendation engine.
That required an invention of new math or new concepts to deliver that. And so it’s an extremely difficult thing to innovate in this space and it’s happening but it takes a lot of resources to do that. And so Amazon and Netflix both who have published papers on how to do it, how do we implement their algorithms.
Amazon has patents on what they’re doing and they’ve they’ve effectively allowed their use outside of Amazon. So those are kind of the two biggies. And they they’ve done a really good job of controlling that back. Also you mentioned Google on the deep learning side.
You know Google’s papers drive a lot of the Internet and tenser flow is huge right now and I just wanted to throw that in there on the deep learning side but we’re not really covering that but they’ve contributed a ton of course they have a ton of resources to put behind that.
Corey: What are some the most common questions that you would hear from a client from, your mom, anyone who might be wondering what is it like what are the questions he might get about machine learning and how to implement it.
Daniel: My mom says Stop stalking me which is okay. I don’t mind that, it actually makes me feel good when she says that because it means we’re doing our job right.
But I mean I think the challenge for marketers and the question we get a lot is how do I how do I trust machine learning, the machines if you want to call it that, to do a good job and provide swim lanes for the machines to stay in bounds.
I think one of the main concerns that we hear repeatedly is that there’s a concern over what the machines are going to do that are kind of outside of brand, outside of brand voice, or brand of messaging or cadence. And the only way we can we can resolve that is by kind of giving the marketer or whoever that that user is kind of controls in those lanes.
And so I think that’s the biggest concern and it’s it’s heard. I think if we just let the machines do their job I don’t know that the outcomes are what we would want them to be. There are some things where it makes sense.
Product recommendations is a great example where clients can hang themselves pretty quickly by putting artificial constraints on what can be recommended and it may be for the purpose of not wanting to sell something that’s going to sell no matter what or because there’s an inventory control issue or something like that.
And those are all fair concerns or fair constraints to put on on a recommender but on some level you want the recommender to do its job because ultimately it’s been tested.
If if the product is worth anything. It’s been tested it’s been weighed against historical data to see if it doing its job accurately and in the end you want to recommend what the person’s going to buy.
Unless you don’t and then you have that’s where you have control. So that’s probably the primary concern. And how how you get around that and on the public side, I think it’s the kind of the stalking aspect of it it feels a little…
I think probably most people have been in a situation where they’ve gone and looked at something on Amazon and then all over the Internet that thing is is on every page everywhere. And there’s going to be a group of people in the world who don’t like that, it just is the nature our human nature to have some privacy.
So there’s a balance that might be a little aggressive. I don’t know it probably works really well if they continue to do it but I think that’s a primary concern. And there are ways to work around that. I think in at least in Cordials case almost everybody everybody who’s getting recommended something as an example has some kind of business relationship with that company.
So it’s not like it’s anonymous data that’s that’s being used to stalk people and and try to get them to become a customer. We’re taking that data and using it in a way that is complementary. So anyways that’s that. I think those are kind of the two classes of concern and they they’re both fair.
Corey: I think it’s important to remember that the customer experience is really kind of the bottom line of what you’re trying to deliver and a bad customer experience is “I feel stocked.” This is weird and I don’t know why I’m seeing the same product that I’ve booked for once everywhere and every page but at the same time you want to deliver that one to one experience where it’s surprising. Right it’s delightful. It’s “wow this is really cool” or “wow I can’t believe that you know I was looking at this thing and now they’re getting an offer for it.” That’s awesome.
Daniel: I think that’s where those controls we talked about come in as you give marketers the ability to put guidelines around the brand and what makes sense for the brand and if you do that ultimately, hopefully your customers don’t feel stocked at that point.
Corey: Yeah absolutely. Let’s go back a little bit. You mentioned that you don’t want them to do things that you don’t want them to do or that they’re bad at. Right. So the machines have rules. Are the machines going to save you time or are you out to manage them. Are they going to cost you a lot of time?
So I guess the answer is kind of twofold. What’s funny is we talk about machines as if they’re like robots that can’t be controlled. And we go back to that.
It’s a complex system it has inputs and outputs and and while there may be some kind of funkiness that goes on inside those algorithms, there are controls around again inputs and outputs so I think that doing it yourself is a challenge.
There are a lot of components in play from infrastructure which is easier with public clouds to kind of data management which probably we find is the most difficult for customers and prospects is dealing with all the data, where they get it, where to put it and when can I get rid of it. And that a lot of questions around that that are that’s probably the most difficult aspect of it.
And then there’s the algorithm component which again the algorithms are only as good as the data you put in. And in the same vein only as good as the data you don’t put in. So sometimes you can put data in that makes the algorithms worse and makes it in a specific case a recommender makes the outcome negative, in terms of its efficacy.
And so managing those algorithms, figuring out the right set of data is key and that becomes a difficult thing and so there are tools to go it alone out there. But again all those three things are keys to making it work. So algorithms, data management, and infrastructure and without the ability to do those on your own, it’s a difficult thing.
So there’s a lot of vendors out there for different things whether it’s subject line text optimization to testing to recommendations and relying on those folks for data management and algorithm side of things is key because there are learnings across all those clients so for example with Cordial we have access to a lot of data across a lot of clients and so we’re able to look at how those algorithms respond in different scenarios before a customer even comes onboard.
Customer comes on board we have already seen similar data because much like humans companies are very similar. Typically they’ll have a product catalog, they’ll have some type of online presence with their pageviews maybe Mobile where we can get mobile data. And we’ve seen that and we know what works and what doesn’t, out of those datasets.
Corey: Last question. Where is it going? Machine learning in five years and two years. Next year. I mean what are some of the innovations that we’re going to see in the market?
Daniel: Yeah I think there’s kind of a couple of things I think know nobody knows kind of where it’s really going to go. Deep learning. Even the guy who developed deep learning doesn’t really know how it works which is makes me chuckle all the time.
But I think deeper things you’ve You mentioned a little bit deep learning is where the algorithms themselves are discovered by the machines, to call them that.
And so there’s even more need for those swimlanes to make sure that kind of everybody stays in order. Deep learning is huge though. I think the idea behind deep learning in terms of marketing is “I’m going to give it all my data and I’m going to get back the best program possible” That might be true. But again it’s still a result of inputs that are provided so deep learning is huge and is going to play a huge role in developing a better experience and I think ultimately driving toward a completely customized, unique experience for every unique person gets their own experience so that’s going to help drive that.
I think the other component is data continues to get cheaper to store.
It continues to be faster to process with better computing power. I think we’ll see a lot of data analysis that happens that gets us to better use of the algorithms that we already have. And so that’s maybe more of a short term thing but using post analysis of what what we recommended and what the outcomes were in doing that on a regular basis.
There are opportunities to do that in a more rapid way and deliver better algorithms or better results from those algorithms faster and so has probably short term, I think deep learning long is a kind of a long term help there.
Corey: Well Daniel thanks for all your wisdom thanks for coming on here. Thanks for your time. You’ve been listening to the Be Cordial podcast. Thanks for listening. And until next time.