Matching AI Models To Business Needs, Supervised Learning Examples — Ad Pricing And Medical Imaging

Managers often have fun when talking to their technical staff. When it comes to artificial intelligence, that fun can have quote marks around it. A few years back, Nvidia CEO Jensen Huang was talking about a “Cambrian Explosion” of artificial intelligence algorithms. While the vast majority of those were in academia, many are making their way to the business world. That might make managers nervous during conversations. Well, relax. It’s not necessary to know about all the different algorithms. On the other hand, it does help to understand the classes of algorithms. The different classes lean towards solutions for different business problems, and it’s good to have a high level understanding of the links.

You may have heard of a number of algorithms (a set of rules that drive software and many other things) mentioned, and wonder why they matter and how they are used. Regression analysis is something that might be remembered from high school or college math. The words “classification” and “clustering” might have been seen, but you wonder why they matter or what’s the difference between them. There are explanations. I and others have provided some in the past, but it’s nice to go through them and link the more technical concepts to the types of business problems then can address.

I recently came upon a book that can help explain the concepts, so I’ll leverage that. “Lean AI”, by Lomit Patel (O’Reilly, ISBN:978-1-492-05931-8), is a book I reviewed elsewhere. Chapter 5 is useful for the current discussion, Figure 5-1 in particular.

I’ve decided that it’s a nice way to cover each type of learning at a high level.

I’ve decided that it’s a nice way to cover each type of learning at a high level.

One caveat, while I am referring to that figure, anything past that is my own opinion and explication, don’t blame Mr. Patel for that.

While it looks complex, there’s a simple explanation. Over the next couple of months, this column will briefly discuss each of the four types of AI learning (the top box in the four groups) and bring that to the business needs. The four primary areas of AI learning are:

·      Supervised: input data is labeled and the output desired is known

·      Unsupervised: non-labeled data and no output specified

·      Semi-supervised: unsurprisingly, a mix of both used when there’s not enough labeled data and when unexpected results might occur

·      Reinforcement learning: trial by error, when the software “experiments” and the trainers give feedback in a cyclical way

In this article, I’m going to address supervised learning. We’ll discuss the types of business problems, then what needs to be done with the data, and then drive to the algorithm that is most useful.

Ad Pricing

The problem of advertisement placement has been around since before Romans had ads written on walls. Newspapers, radio and television expanded the number of places and costs of ad options. Now we add the internet, with web sites, social media platforms and more. How is a marketing group to figure out the best places for ads, especially as prices on the internet can change minute by minute?

A marketing group, or more likely a third party provider, can track an enormous amount of data. For this example, think of all the vectors that can be tracked. They include, in part: price of ads, length of ad, length of placement, number of viewers, and geographic spread. How to balance those areas requires looking at how they relate and choosing what is hoped to be the best price-performance mix.

Think back to High School algebra. Linear regression, in the simplest form, relates two variables to each other. For instance, Figure 2 might compare cost and length of placement for ads at different outlets. With the number of variables above, and the large amount of data, regression becomes far more complex.

When it comes to a machine learning (ML) system, when we know exactly where all the data comes from and which data is important, it can be accurately labeled. That’s when supervised learning is the option.Medical Imaging

One of the first areas of business focused on by AI teams has been medical imaging. Radiology is critical to finding and identifying cancers, tumors, and other medical problems. Doctors and technicians have to look at many images and find discrepancies that can be difficult for the human eye to identify.

What the people do is look at images of tissues and then identify good tissue and bad tissue. That is a problem called classification. The categories to be identified are understood, the goal is to put images and sections of images into each category.

Classification, because of the known categories, falls into the supervised learning method for training ML systems. Good and bad tissue can be identified, even by the type of tissue, so the data can be clearly labeled. The output can then point to the resulting classes the image falls into.

Supervised Learning: When You Know What You Want

Supervised learning is appropriate for business problems where you know the results. When features can be clearly defined and labeled, the ML system can be trained on the expected results. When you are looking at a problem, such as home prices based on features or visually sorting inventory, labeling leads to supervised learning training.