What Is Machine Learning?

After stints at MIT and Bell Telephone Laboratories, Arthur L. Samuel joined IBM in 1949 at the Poughkeepsie Laboratory. His efforts helped boost the computing power of the company’s machines, such as with the development of the 701 (this was IBM’s first commercialized computer system).

But he also programmed applications. And there was one that would make history—that is, his computer checkers game. It was the first example of a machine learning system (Samuel published an influential paper on this in 1959⁶). IBM CEO Thomas J. Watson, Sr., said that the innovation would add 15 points to the stock price!⁷

Then why was Samuel’s paper so consequential? By looking at checkers, he showed how machine learning works—in other words, a computer could learn and improve by processing data without having to be explicitly programmed. This was possible by leveraging advanced concepts of statistics, especially with probability analysis. Thus, a computer could be trained to make accurate predictions.

This was revolutionary as software development, at this time, was mostly about a list of commands that followed a workflow of logic.

To get a sense of how machine learning works, let’s use an example from the HBO TV comedy show Silicon Valley. Engineer Jian-Yang was supposed to create a Shazam for food. To train the app, he had to provide a massive dataset of food pictures. Unfortunately, because of time constraints, the app only learned how to identify…hot dogs. In other words, if you used the app, it would only respond with “hot dog” and “not hot dog.”

While humorous, the episode did a pretty good job of demonstrating machine learning. In essence, it is a process of taking in labeled data and finding relationships. If you train the system with hot dogs—such as thousands of images—it will get better and better at recognizing them.

Yes, even TV shows can teach valuable lessons about AI!

But of course, you still need much more. In the next section of the chapter, we’ll take a deeper look at the core statistics you need to know about machine learning. This includes the standard deviation, the normal distribution, Bayes’ theorem, correlation, and feature extraction.

Then we’ll cover topics like the use cases for machine learning, the general process, and the common algorithms.

Standard Deviation

The standard deviation measures the average distance from the mean. In fact, there is no need to learn how to calculate this (the process involves multiple steps) since Excel or other software can do this for you easily.

To understand the standard deviation, let’s take an example of the home values in your neighborhood. Suppose that the average is $145,000 and the standard deviation is $24,000. This means that one standard deviation below the average would be $133,000 ($145,000 – $12,000) and one standard deviation above the mean would come to $157,000 ($145,000 + $12,000). This gives us a way to quantify the variation in the data. That is, there is a spread of $24,000 from the average.

Next, let’s take a look at the data if, well, Mark Zuckerberg moves into your neighborhood and, as a result, the average jumps to $850,000 and the standard deviation is $175,000. But do these statistical metrics reflect the valuations? Not really. Zuckerberg’s purchase is an outlier. In this situation, the best approach may be instead to exclude his home.

The Normal Distribution

When plotted on a graph, the normal distribution looks like a bell (this is why another name for it is the “bell curve”). It represents the sum of probabilities for a variable. Interestingly enough, the normal curve is common in the natural world, as it reflects distributions of such things like height and weight.

A general approach when interpreting a normal distribution is to use the 68-95-99.7 rule. This estimates that 68% of the data items will fall within one standard deviation, 95% within two standard deviations, and 99.7% within three standard deviations.

A way to understand this is to use IQ scores. Suppose the mean score is 100 and the standard deviation is 15. We’d have this for the three standard deviations, as shown in Figure 3-1.

Note that the peak in this graph is the average. So, if a person has an IQ of 145, then only 0.15% will have a higher score.

Now the curve may have different shapes, depending on the variation in the data. For example, if our IQ data has a large number of geniuses, then the distribution will skew to the right.

Bayes’ Theorem

As the name implies, descriptive statistics provides information about your data. We’ve already seen this with such things as averages and standard deviations.

But of course, you can go well beyond this—basically, by using the Bayes’ theorem. This approach is common in analyzing medical diseases, in which cause and effect are key—say for FDA (Federal Drug Administration) trials.

To understand how Bayes’ theorem works, let’s take an example. A researcher comes up with a test for a certain type of cancer, and it has proven to be accurate 80% of the time. This is known as a true positive.

But 9.6% of the time, the test will identify the person as having the cancer even though he or she does not have it, which is known as a false positive. Keep in mind that—in some drug tests—this percentage may be higher than the accuracy rate!

And finally, 1% of the population has the cancer.

Sounds kind of out of whack, right? Definitely. After all, how is it that a test, which is 90% accurate, has only a 7.8% probability of being right? But remember the accuracy rate is based on the measure of those who have the flu. And this is a small number since only 1% of the population has the flu. What’s more, the test is still giving off false positives. So Bayes’ theorem is a way to provide a better understanding of results—which is critical for systems like AI.

Correlation

A machine learning algorithm often involves some type of correlation among the data. A quantitative way to describe this is to use the Pearson correlation, which shows the strength of the relationship between two variables that range from 1 to –1 (this is the coefficient).

Then what is a strong correlation? As a general rule of thumb, it’s if the coefficient is +0.7 or so. And if it is under 0.3, then the correlation is tenuous.

All this harkens the old saying of “Correlation is not necessarily causation.” Yet when it comes to machine learning, this concept can easily be ignored and lead to misleading results.

There is a name for this: patternicity. This is the tendency to find patterns in meaningless noise.

Feature Extraction

In Chapter 2, we looked at selecting the variables for a model. The process is often called feature extraction or feature engineering.

An example of this would be a computer model that identifies a male or female from a photo. For humans, this is fairly easy and quick. It’s something that is intuitive. But if someone asked you to describe the differences, would you be able to? For most people, it would be a difficult task. However, if we want to build an effective machine learning model, we need to get feature extraction right—and this can be subjective.

This just scratches the surface as I’m sure you have your own ideas or approaches. And this is normal. But this is also why such things as facial recognition are highly complex and subject to error.

Feature extraction also has some nuanced issues. One is the potential for bias. For example, do you have preconceptions of what a man or woman looks like? If so, this can result in models that give wrong results.

Because of all this, it’s a good idea to have a group of experts who can determine the right features. And if the feature engineering proves too complex, then machine learning is probably not a good option.

But there is another approach to consider: deep learning. This involves sophisticated models that find features in a data. Actually, this is one of the reasons that deep learning has been a major breakthrough in AI. We’ll learn more about this in the next chapter.

What Can You Do with Machine Learning?

As machine learning has been around for decades, there have been many uses for this powerful technology. It also helps that there are clear benefits, in terms of cost savings, revenue opportunities, and risk monitoring.

Figure 3-2 shows some of the applications for machine learning.

The Machine Learning Process

To be successful with applying machine learning to a problem, it’s important to take a systematic approach. If not, the results could be way off base.

First of all, you need to go through a data process, which we covered in the prior chapter. When this is finished, it’s a good idea to do a visualization of the data. Is it mostly scattered? Or are there some patterns? If the answer is yes, then the data could be a good candidate for machine learning.

The goal of the machine learning process is to create a model, which is based on one or more algorithms. We develop this by training it. The goal is that the model should provide a high-degree of predictability.

Now let’s take a closer look at this (by the way, this will also be applicable for deep learning, which we’ll cover in the next chapter):

Applying Algorithms

Some algorithms are quite easy to calculate, while others require complex steps and mathematics. The good news is that you usually do not have to compute an algorithm because there are a variety of languages like Python and R that make the process straightforward.

As for machine learning, an algorithm is typically different from a traditional one. The reason is that the first step is to process data—and then, the computer will start to learn.

Even though there are hundreds of machine learning algorithms available, they can actually be divided into four major categories: supervised learning, unsupervised learning, reinforcement learning, and semi-supervised learning. We’ll take a look at each.

Common Types of Machine Learning Algorithms

There is simply not enough room in this book to cover all the machine learning algorithms! Instead, it’s better to focus on the most common ones.

Figure 3-3 shows a general framework for machine learning algorithms.

Naïve Bayes Classifier (Supervised Learning/Classification)

Earlier in this chapter, we looked at Bayes’ theorem. As for machine learning, this has been modified into something called the Naïve Bayes Classifier. It is “naïve” because the assumption is that the variables are independent from each other—that is, the occurrence of one variable has nothing to do with the others. True, this may seem like a drawback. But the fact is that the Naïve Bayes Classifier has proven to be quite effective and fast to develop.

There is another assumption to note as well: the a priori assumption. This says that the predictions will be wrong if the data has changed.

A common use case for Naïve Bayes Classifiers is text analysis. Examples include email spam detection, customer segmentation, sentiment analysis , medical diagnosis, and weather predictions. The reason is that this approach is useful in classifying data based on key features and patterns.

To see how this is done, let’s take an example: Suppose you run an e-commerce site and have a large database of customer transactions. You want to see how variables like product review ratings, discounts, and time of year impact sales.

Using this chart, we can see that the probability of a purchase when there is a low product review is 7/24 or 29%. In other words, the Naïve Bayes Classifier allows more granular predictions within a dataset. It is also relatively easy to train and can work well with small datasets.

K-Nearest Neighbor (Supervised Learning/Classification)

The k-Nearest Neighbor (k-NN ) is a method for classifying a dataset (k represents the number of neighbors). The theory is that those values that are close together are likely to be good predictors for a model. Think of it as “Birds of a feather flock together.”

A use case for k-NN is the credit score, which is based on a variety of factors like income, payment histories, location, home ownership, and so on. The algorithm will divide the dataset into different segments of customers. Then, when there is a new customer added to the base, you will see what cluster he or she falls into—and this will be the credit score.

K-NN is actually simple to calculate. In fact, it is called lazy learning because there is no training process with the data.

To use k-NN, you need to come up with the distance between the nearest values. If the values are numerical, it could be based on a Euclidian distance, which involves complicated math. Or, if there is categorical data, then you can use an overlap metric (this is where the data is the same or very similar).

Next, you’ll need to identify the number of neighbors. While having more will smooth the model, it can also mean a need for huge amount of computational resources. To manage this, you can assign higher weights to data that are closer to their neighbors.

Linear Regression (Supervised Learning/Regression)

Linear regression shows the relationship between certain variables. The equation—assuming there is enough quality data—can help predict outcomes based on inputs.

As you can see, the general relationship is positive (this describes the tendency where a higher grade is correlated with more hours of study). With the regression algorithm, we can plot a line that has the best fit (this is done by using a calculation called “least squares,” which minimizes the errors). See Figure 3-4.

From this, we get the following equation:

Grade = Number of hours of study × 0.03731 + 0.6889

Then, let’s suppose you study 4 hours for the exam . What will be your estimated grade? The equation tells us how:

0.838 = 4 × 0.03731 + 0.6889

How accurate is this? To help answer this question, we can use a calculation called R-squared. In our case, it is 0.9180 (this ranges from 0 to 1). The closer the value is to 1, the better the fit. So 0.9180 is quite high. It means that the hours of study explains 91.8% of the grade on the exam.

Now it’s true that this model is simplistic. To better reflect reality, you can add more variables to explain the grade on the exam—say the student’s attendance. When doing this, you will use something called multivariate regression .

Sometimes data may not be in a straight line either, in which case the regression algorithm will not work . But you can use a more complex version, called polynomial regression.

Decision Tree (Supervised Learning/Regression)

No doubt, clustering may not work on some datasets. But the good news is that there are alternatives, such as a decision tree. This approach generally works better with nonnumerical data.

The start of a decision tree is the root node, which is at the top of the flow chart. From this point, there will be a tree of decision paths, which are called splits. At these points, you will use an algorithm to make a decision, and there will be a probability computed. At the end of the tree will be the leaf (or the outcome).

A famous example—in machine learning circles —is to use a decision tree for the tragic sinking of the Titanic. The model predicts the survival of a passenger based on three features: sex, age, and the number of spouses or children along (sibsp). Here’s how it looks, in Figure 3-5.

There are clear advantages for decision trees. They are easy to understand, work well with large datasets, and provide transparency with the model.

However, decision trees also have drawbacks. One is error propagation. If one of the splits turns out to be wrong, then this error can cascade throughout the rest of the model!

Next, as the decision trees grow, there will be more complexity as there will be a large number of algorithms. This could ultimately result in lower performance for the model .

Ensemble Modelling (Supervised Learning/Regression)

Ensemble modelling means using more than one model for your predictions. Even though this increases the complexity, this approach has been shown to generate strong results.

To see this in action, take a look at the “Netflix Prize,” which began in 2006. The company announced it would pay $1 million to anyone or any team that could improve the accuracy of its movie recommendation system by 10% or more. Netflix also provided a dataset of over 100 million ratings of 17,770 movies from 480,189 users.¹⁶ There would ultimately be more than 30,000 downloads.

Why did Netflix do all this? A big reason is that the company’s own engineers were having trouble making progress. Then why not give it to the crowd to figure out? It turned out to be quite ingenious—and the $1 million payout was really modest compared to the potential benefits.

The contest certainly stirred up a lot of activity from coders and data scientists, ranging from students to employees at companies like AT&T.

Netflix also made the contest simple. The main requirement was that the teams had to disclose their methods , which helped boost the results (there was even a dashboard with rankings of the teams).

But it was not until 2009 that a team—BellKor’s Pragmatic Chaos—won the prize. Then again, there were considerable challenges.

So how did the winning team pull it off? The first step was to create a baseline model that smoothed out the tricky issues with the data. For example, some movies only had a handful of ratings, whereas others had thousands. Then there was the thorny problem where there were users who would always rate a movie with one star. To deal with these matters, BellKor used machine learning to predict ratings in order to fill the gaps.

The winning team used ensemble modelling, which involved hundreds of algorithms. They also used something called boosting, which is where you build consecutive models. With this, the weights in the algorithms are adjusted based on the results of the previous model, which help the predictions get better over time (another approach, called bagging, is when you build different models in parallel and then select the best one).

But in the end, BellKor found the solutions. However, despite this, Netflix did not use the model! Now it’s not clear why this was the case. Perhaps it was that Netflix was moving away from five-star ratings anyway and was more focused on streaming. The contest also had blowback from people who thought there may have been privacy violations.

Regardless, the contest did highlight the power of machine learning—and the importance of collaboration.

K-Means Clustering (Unsupervised/Clustering)

The k-Means clustering algorithm, which is effective for large datasets, puts similar, unlabeled data into different groups. The first step is to select k, which is the number of clusters. To help with this, you can perform visualizations of that data to see if there are noticeable grouping areas.

Here’s a look at sample data, in Figure 3-6:

For this example, we assume there will be two clusters, and this means there will also be two centroids. A centroid is the midpoint of a cluster. We will assign each randomly, which you can see in Figure 3-7.

As you can see, the centroid at the top left looks way off, but the one on the right side is better. The k-Means algorithm will then calculate the average distances of the centroids and then change their locations . This will be iterated until the errors are fairly minimal—a point that is called convergence, which you can see with Figure 3-8.

Granted, this is a simple illustration. But of course, with a complex dataset, it will be difficult to come up with the number of initial clusters. In this situation, you can experiment with different k values and then measure the average distances. By doing this multiple times, there should be more accuracy.

Then why not just have a high number for k? You can certainly do this. But when you compute the average, you’ll notice that there will be only incremental improvements. So one method is to stop at the point where this starts to occur. This is seen in Figure 3-9.

However, k-Means has its drawbacks . For instance, it does not work well with nonspherical data, which is the case with Figure 3-10.

With this, the k-Means algorithm would likely not pick up on the surrounding data, even though it has a pattern. But there are some algorithms that can help, such as DBScan (density-based spatial clustering of applications with noise), which is meant to handle a mix of widely varying sizes of datasets. Although, DBScan can require lots of computational power.

Next, there is the situation where there are some clusters with lots of data and others with little. What might happen? There is a chance that the k-Means algorithm will not pick up on the light one. This is the case with Figure 3-11.

Conclusion

These algorithms can get complicated and do require strong technical skills. But it is important to not get too bogged down in the technology. After all, the focus is to find ways to use machine learning to accomplish clear objectives.

Again, Stich Fix is a good place to get guidance on this. In the November issue of the Harvard Business Review, the company’s chief algorithms officer, Eric Colson, published an article, “Curiosity-Driven Data Science.”¹⁷ In it, he provided his experiences in creating a data-driven organization.

At the heart of this is allowing data scientists to explore new ideas, concepts, and approaches. This has resulted in AI being implemented across core functions of the business like inventory management, relationship management, logistics, and merchandise buying. It has been transformative, making the organization more agile and streamlined. Colson also believes it has provided “a protective barrier against competition.”

Key Takeaways