You Ask, I Answer: Classes of Problems AI Solves in Marketing?

In today's episode, Annika asks: important functions of AI with big data are analyzing the past, predicting the future, and prescribing future strategies according to the Marketing AI Institute. What else do you know to be important use cases for AI and marketing? So this is one of the things, the most important things we need to understand about AI is that it isn't just a collection of random solutions and operations, right? Uh what we have happening here is an attempt to try to understand what AI can do and and conflate it with business outcomes. And those are two different things.

So what we need to do is dig into the operations, the capabilities of AI, machine learning particularly, and understand that framework first, and that it's not random and that's not uh just a collection of of disconnected point solutions. Understand the framework, the structure first, and then apply it to um the business outcomes we care about. So let's look at what this framework looks like. In fact, let's bring it up here. So artificial intelligence, machine learning in particular, is good at four types of problems, right?

In the upper left, you have well, you have two types of out of problems you're trying to solve. You either know the outcome you're looking for, which is called supervised learning, or you have you don't know what you're looking for. You have a big pile of data and you're trying to make sense of it. That's called unsupervised learning. So that's the top.

Along the left hand side, you have two types of data. You have numbers, continuous data, metrics, and you have non-numbers. So stuff that's categorical in nature, dimensions, descriptors, things like that. An example continuous data of course is any kind of number. Categorical data would be things like place names, um channels in Google Analytics, right?

Facebook, uh so you see this grid supervised, unsupervised, continuous categorical. In each of these four categories, uh there are different types of uh techniques. So it for when you know the outcome you're after and you have numbers you can do regression and prediction. So regression very straightforward mathematical operations right prediction very similar um that's where you get things like gradient boosting. That's where you get things like um uh GBM XG boost uh even just you know good old fashioned uh regular linear regression all that is to get to find an um to find the outcome you're looking for right so you want to know for example uh what drives website traffic or what drives conversions it's regression and prediction uh time series forecasting is also in this bucket as well because it's a form of regression when you have unsupervised uh machine learning and continuous data you have clustering so you have a bunch of numbers.

How do they cluster together? How do they relate to each other um this is a great way to dig into things like um for example, SEO data. When you have a lot of SEO data uh you have a lot of different numbers. Domain authority, link authority, inbound links, outbound links, number of words, etc. They're all numbers.

How do they relate to each other? You need to cluster it first to make start making sense of the data, and then only then would you then flip back over to regression and say, okay, now that I've made sense of the data, which of these clusters has a relationship to high search rankings? So clustering is is a very important set of techniques. We use it a lot. Classification and categorization and categorical supervised.

This is all speech recognition, uh, image recognition, video recognition. This is uh classifying uh sentiment analysis, things like that. Anything that is non numerical, you're gonna be doing a ton of this classifying. Um Bayesian learning and stuff is all in here. Uh logistic regression is in here, and those, of course, their obvious applications, speech recognition, natural language processing, and so on and so forth, um, are all in the classification bucket.

And then association and dimension reduction, that's where you get a lot of uh natural language understanding. So trying to understand the words, the phrases, the uh bigrams, trigrams uh in text. Um you see a lot of that uh you see it in in the first stages of image recognition as well, before you go to classification, just trying to pick out the noise from the image. What's foreground, what's background? When you understand this framework of the types of machine learning, uh then you can start mapping it to those business outcomes.

But you have to understand this first. If you don't have this stuff down first and understand the techniques that go in each of these buckets, it becomes very, very difficult to make sense of AI. Because what happens, of course, is that you it just gets too messy, right? And you don't if you don't have this in your head, then you don't have the underlying techniques that go in it. So in supervised, you have, for example, ordinal regression, random forest regression, boosted tree regression, linear regression, and unsupervised, you have K-means clustering, uh, covariance clustering.

Uh for categorical stuff, you have decision trees, logistic regression, neural networks, nearest neighbor. I talked about I mentioned Bayesian early. Uh for association dementia reduction, you have stuff like PCA, uh, LDA, CCA, TSNE. Um, but if you don't understand these things and you don't understand the techniques inside of each of them, then it becomes very difficult to apply them correctly. So understand this stuff first and the techniques within them, and then start thinking about okay, now how do I apply this to business outcomes?

If you don't have the command of the techniques, then one of the consequences of this is that uh it becomes very easy for vendors to fool you. It becomes very, very easy for vendors to take advantage of you to say, oh, yeah, we use machine learning, we use uh can you imagine a vendor saying we used advanced uh uh machine learning and and proprietary dimension reduction techniques to predict the future? Well, no. You may do dimension reduction to clean up the data, but if you're predicting f using dimension reduction techniques, you're literally doing th completely opposite of the way they should be. So this is it's important for marketers to have this vocabulary.

Um you don't necessarily need to write the code, but you need to understand the techniques involved in order to map them to use cases. Once you understand the data each works with and the techniques in each, then finding use cases for all this stuff is very, very straightforward. Not easy, but straightforward. So an important question because understand this stuff first and then go hunting for use cases for AI and machine learning. A lot to unpack here, a lot to study.

I'm still learning. Uh, everybody I know in the field is still learning. Uh, there's new techniques being developed all the time or improvements to techniques, so keep that in mind. As always, please leave your comments below. Subscribe to the YouTube channel to the newsletter.

I'll talk to you soon. Want help solving your company's data analytics and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.