The marketing technology stack is accumulating more and more products that claim to have artificial intelligence. What does that mean, and should you care? How can AI help your business?
At this stage of the game, I think there are three important things to focus on.
- There is simply too much data out there for humans to analyze effectively, but it needs to be analyzed. We need the help of smart systems to find patterns buried in the masses of data we collect.
- Business people need to know enough about artificial intelligence and machine learning to avoid being deceived by false promises. AI can sound too much like a system that's going to magically solve your problems. In reality, it often involves a lot of work.
- Business people need to have a high-level knowledge of the types of machine learning algorithms that would be helpful with their data.
Warning: I'd like to clarify that I'm not an expert on this topic. I've done research for this article, and I’ve run it by a couple friends who know the subject better than I do, but don't take what I say as Gospel. If you see something amiss, please point it out.
The way some people use the terms, artificial intelligence is the umbrella term, while machine learning is a subset.
I think it's a little more helpful to look at the distinction this way: machine learning is about pattern recognition and making predictions, while artificial intelligence is about making decisions and taking actions based on those predictions. As an example, a machine learning routine in a self-driving car might predict that the jaywalking pedestrian is a risk, but the artificial intelligence routine applies the brakes.
Creative and abstract thinking would also be artificial intelligence, but I don't think anybody's been able to get a computer to do that yet.
Machine learning is the process by which a program can learn from its experiences. The program uses statistics to find and learn patterns.
Pro tip: Some machine learning algorithms require human supervision, while others are unsupervised. That’s a key distinction you’ll need to pay attention to. Be sure to ask the sales team about that.
Popular business uses for supervised machine learning include applications like personalized marketing, fraud detection, people analytics, and insurance underwriting decisions. Unsupervised uses often focus on customer grouping, anomaly detection, and product affinity/association rule engines.
An example of supervised machine learning might be a program that's designed to distinguish different types of hats — say, to group pictures of fedoras and pictures of bowlers. The program might have several different variables, like the size of the brim relative to the hat, the curvature of the brim, the shape of the top of the hat, and so on. The program is then fed a collection of test data and a human checks the program's results to see if the program got it right. The program then adjusts how it uses the various parameters until its success rate is acceptable.
Note that supervised machine learning is not a matter of plug it in, turn it on and magical results come out. There's a lot of training and fine-tuning involved. From a nuts and bolts standpoint, supervised learning typically involves classification and regression techniques (as discussed below) while unsupervised learning leans on clustering algorithms.
If you're considering a new technology that includes machine learning, ask the sales team whether it's supervised or unsupervised machine learning. If it's supervised, you need to dig in to find out how much work you're setting yourself up for. The lion’s share of this supervised work typically involves applying classification techniques based on labeled data. There is no magic here, input data needs to be proactively labeled by humans at some point, E.g., this is a duck, this is a rabbit, etc. The classification burden grows as a function of data size and complexity.
It's also important to note that no matter how good that hat-discriminating program becomes, it probably won’t be able to distinguish a cowboy hat from a baseball hat, because it's been trained on fedoras and bowlers. It might be very good in its own, narrow scope, but it fails when it's applied beyond that scope.
Pro tip: Ask the sales team how narrowly or broadly an algorithm can be applied. What are its strengths and weaknesses?
Don't think machine learning is like Commander Data from Star Trek, or Hal from 2001. It can't extrapolate or engage in creative thinking. If someone creates a program that's great at playing Go on a standard board, that same program would fail if you asked it to play on a larger board. A human could extrapolate from the one set of rules to the other, but the computer can't. At least not yet.
As a not entirely precise rule of thumb, machine learning is about pattern recognition. AI is about being intelligent with the results of machine learning — to make decisions based on those predictions.
Different types of algorithms are useful in different situations, and there are a few different ways of categorizing them. My focus recently has been on Customer Data Platforms. In that context, these seem to be the most important types of machine-learning algorithms. My uber-geek friends say I’ve missed a few, but I don’t want this to be overwhelming.
Supervised ML Approaches
Convolutional Neural Networks are most commonly applied to visual imagery and recommendation systems. They can be used for both classification and regression.
Classification techniques are used to predict categorical values, such as whether a hat is a bowler or a fedora. Here are two types of classification techniques.
- K nearest neighbor defines clusters, where the user determines how many clusters are in the data. The algorithm groups data based on a set of features. E.g., is this song country, bluegrass or rock and roll?
- Decision trees create simple rules from the training data to predict possible consequences of particular actions. E.g., if this patient has these symptoms, will this medicine help?
Regression is used to predict continuous numerical values.
- Logistic regression is used to estimate the probability of a binary response, e.g., is the email spam or not?
- Linear regression models a relationship between observation and outcome using a straight line, such as spend on advertising vs. revenue.
- Polynomial regression looks for relationships that exist on a curve — for example, a square relationship.
Unsupervised ML Approaches
Collaborative Filtering is what lies behind many recommendation engines. The assumption is that if person A is like person B in one set of characteristics (e.g., musical preferences), person A is more likely than a random person to be like person B in other characteristics.
Clustering analysis is the most common unsupervised learning method. It's used to find hidden patterns or groupings of data.
- K-means clustering partitions data into clusters based on the nearest average. I find this one is easiest to understand if you watch it happen.
- Principal component analysis tries to find the best-fitting line(s) that predict patterns in a dataset. For example, what characteristics of a house are mostly likely to correlate with a high selling price?
Summary
Under supervised learning, input data is labeled up front with finite classifications, while the same is often not true with unsupervised learning. Supervised machine learning tends to be more computationally complicated than unsupervised machine learning, but it typically yields more accurate results.
“When you evaluate the various applications of machine learning, it is not always easy to choose which solutions would be the most feasible and effective for your business. The technique that you choose will depend on your business’ goals and objectives, availability of resources and access to data.” (Source.)
There are other types of machine-learning algorithms out there, and both YouTube and Wikipedia have lots of good information on this subject. I hope my brief introduction helps you navigate through it all.
Today's marketers are collecting extraordinary amounts of data on their customers. It's no longer possible to toss it all into Excel and make a rational business decision. Some level of machine learning is necessary, and Customer Data Platforms are a good place for this to happen.
If you have questions about this article, or about CDPs, give me a shout.
If you liked this article, please send it to a friend, or maybe post it to LinkedIn.
Great article, Greg! This makes the differences between AI and ML easy to understand for those of us who are still figuring it out. You lost me at “Convolutional Neural Networks,” but now I have a good basic understanding. I’ll definitely be sharing this article with others!