An inspectable algorithm allows you to review the results and understand what the model is doing. Decision trees are the only algorithm that enables a user to inspect and gain insights into the results produced by the algorithm. Decision trees have existed since the 1960s and continue to be used. However, they have a major limitation: Decision trees cannot support a large number of features or large data sets. Because of these scalability limitations, it can be difficult to ascertain the patterns that are at the heart of machine learning for large data sets. Despite the fact that inspectability or explainability is instrumental in successfully understanding what models are doing, decision trees serve as only one type of tool in the machine learning predictive toolbox.

Now let’s look a little closer at how decision trees work. By understanding them, you will recognize the value of a ML algorithm. Looking a little closer at how this algorithm functions, you will see that it captures relationships among categories by making decisions as it branches down the tree on what feature matters in the decision in the tree. The “leaf” or ends of the tree represent the categories.

Let’s start with a hypothetical example. Suppose that a tire company wants to know the causes of tire failures at less than 6 months of age of their tires. The tire engineers must first create the data set for the algorithm to use. This process is not trivial and takes some guess work and a lot of data modeling to get the data ready to be used. The tire engineers guess that important features of the data could include the type of tire (model 01, 02, 03), the chemical types (assume these are rubber and synthetic nylon, for this discussion), the tire weight, and the length of time it was stored. Scientific articles suggest that the day of the week might also be a feature, because workers tend to be less productive and mindful on Mondays. The engineers decide to use as features for storage amounts of less than 5 days and equal to or more than 5 days. The same algorithm could be run with different time features to see what results occur. Although the discussion below uses terms such as rubber, synthetic, and days of the week, in fact those values would need to be turned into numerical forms (1 or 2 for tire type; 1–7 for days of the week, and so on) for processing in the algorithm.

The algorithm uses these features for each example in the data set it learns from. It creates the decision tree as it learns from examples in the data set provided by the company. The hypothetical resulting trees would look like the ones in Figure 4-1 on the next page.

Inspecting these results, one sees that (1) tire weight and model did not matter, as they are not factors in the decision trees; (2) overall rubber tires had higher failure rates than synthetics; and (3) storage and creation on Monday for both categories of tires were a problem but worse for rubber. The highest failure rates were for rubber tires stored more than 5 days and created on a Monday. There is no guarantee that using a larger data set with the same features would result in the same tree, but unless the smaller data set is anomalous in its examples, the larger data set might have slightly different results but not be fundamentally different. Checking that assumption is very important in using machine learning algorithms as well as trying different features of the tires. This decision tree tells the engineers what kind of failure rates they can expect for their tires depending on when they were made. However, over time, if the engineers improve their tire manufacturing, a new data set might show different results from the one in this example. So it is critical to keep collecting new data and seeing what results occur for that data.

Note that, importantly, decision tree algorithms are not widely used when there are a large number of features or when the data set is quite large because the algorithm runs too slowly to produce the model. So why use decision trees at all? In inspecting the decision tree, developers and business users can explain just how it is that a new tire will be categorized. Imagine if the tree was not inspectable. We would place a black piece of paper on the whole tree in the diagram and just leave the bottom leaves of the tree to look at. Then, when a given tire was run through the model, the results would simply show that it had one of the failure rates at the leaves of the tree. No one would know why that tire had that failure rate, or what features contribute to that failure rate, and they could not use the algorithm to decide what to change in their business process! Obviously knowing the predicted failure rate is useful, but knowing why is even more so. Furthermore, inspecting the tree may also help the human team see if there is bias in the original data, a matter we will elaborate upon in Chapter 5.

The vast majority of machine learning algorithms provide non-inspectable results. The algorithms are often called black box algorithms because unlike a decision tree, you can’t look inside to see how the algorithm works. Non-inspection means that the way that the algorithm makes its choices is not something that a data scientist or anyone else can look at, just like when we put black paper over the decision tree. As a result, black box algorithms require other methods to see whether they are producing reasonable results.

Opaque or black box algorithms do not provide the analyst with an understanding of why the algorithm produced its results. The algorithm makes choices and reaches conclusions based on how it interprets the data. Therefore, if an organization needs to explain how a conclusion was arrived at, it will require different methods to explain the results. For example, if a machine learning algorithm is used to make hiring recommendations, an organization may be required to explain why certain individuals were hired and why others were not considered.

There are two types of opaque algorithms: supervised and unsupervised. As we discussed in Chapter 2, a supervised model is one in which you begin with an established set of data that is already classified. This classified data has labeled features that define the meaning of the data. In contrast, unsupervised models are based on a large volume of unlabeled data that must be classified based on a pattern or cluster.

All supervised algorithms start with examples for which the data scientist knows how the classification should go (sometimes called the gold standard). The algorithm must learn this classification. However, because the rules the algorithm learns cannot be inspected, some of the examples must be set aside for later evaluation (called the test set or test data). The standard protocol for these algorithms is to have the algorithm train on the majority of the examples, called the training set or training data, which is often 90% of the examples. The remaining 10% of the data are the test set. The algorithm is never trained using the test set because it would eliminate the method of seeing how the algorithm performs on “new” data. When the algorithm finishes on the training set, it is then evaluated by testing it using a test set to see how well it performs.

K-nearest neighbors works by taking an example’s value and then locating another example value that is close to it (thus the term nearest neighbor). What counts as close in distance depends on the particular features that the data scientist chooses. For the tire example, these features could be age (in whole numbers of 1 to 120), weight (in whole numbers or real numbers), and so on. This algorithm can create many clusters (not just failure or no failure for the tire business).

An important challenge for this algorithm is the size of k. The k in a k-NN algorithm stands for a number that is a positive integer that must be chosen when running the algorithm. If k is too small, the algorithm will overfit the data, which means it will fail to predict certain members of the clusters. Splitting into a training and test set, and keeping the test set large enough, helps to reduce overfitting. If k is too large, irrelevant examples will creep into clusters and make the algorithm unreliable. Irrelevant examples are referred to as outliers. Data scientists sometimes have ways of visualizing data (with graphs) that allow them to remove outliers from the data set before the algorithm works, but data sets with many features do not make this visualization straightforward. So data experts often try several different values of k in order to see how well the algorithm performs on the training set and the test set. Of course, test sets are never used for training data.

A support vector machine is widely used when there are a small number of features to train on, where “small” depends on the number of features relative to the number of items in the data set. The tire example discussed so far has a much smaller number of features relative to the typical data set in most companies. A support vector machine makes its decisions not by simply drawing a line between the members of the data set. Instead, it chooses a margin around that line and looks to make the margin as wide as possible.

There are many kinds of support vector machine algorithms, which grow in complexity with the mathematics that govern them (from linear SVMs to ones with quadratic functions and beyond). Most data scientists prefer linear SVMs or ones that use quadratics because they are simpler to compute. SVMs have been used to classify texts (to determine which of a set of prespecified categories a text most closely fits), perform facial recognition (to find the face or faces in an image), image classification (again to pick the prespecified category in which the image fits), interpret handwritten letters, and to classify proteins. Note that in all of these applications, a set of classes is prespecified, not invented by the algorithm. Like other supervised machine learning algorithms, SVMs rely on people to provide the shape of the problem that the machine algorithm is to learn, and the algorithm provides the (non-inspectable) rules to predict which category an example will fit into.

Here we take a moment to consider a business problem application so that we can understand how algorithms might work with it. A cable television company that provides video content to viewers wants to know what advertisements to show to its customer viewers while they watch a show or movie. The data that the company has will include customer location, the show being watched, the history of other shows watched, the time of day (morning, afternoon, early evening, late evening, past-midnight), and the content package the customer pays for (for example, bare bones, many channels, special channels, etc.). The company may purchase information about the income level of the customer (such data is readily available), but it may still not have certain types of significant data, such as the gender and age of the viewer, and whether a group or a single individual is watching. This missing data might be crucial to the model and is a potential weakness in the model.

The company must determine what categories of advertisements it wants the algorithm to display using the data set and the categories of location, show, and so on. For example, does the company have six different categories of advertisements it wants to consider or does it not know what categories of advertisements it should use? When the company knows the categories of advertisements it wants to categorize its data into, it can choose a supervised algorithm. The algorithm will create a model that has learned a way to decide the categories of advertisements to show viewers as they watch a particular show. For example, the model might have learned that evening viewers watching shows about international travel should be shown advertisements about cars and household appliances. However, the cable company must also assure that both training and test sets contain representative data for each of the categories that the algorithm will learn from or else the learned model will make poor predictions for some viewers. If the company does not have a way to create such sets, it is best to turn to unsupervised algorithms. Furthermore, if the company does not know what categories it wants, it can use unsupervised algorithms to determine the categories and put values into each.

Unsupervised learning algorithms for clustering tasks fall into two broad classes: K-means clustering and hierarchical clustering. There are other types of tasks for unsupervised learning to reduce the dimensionality of a problem, but such tasks are often used by data scientists in order to simplify problems. K-means clustering is one of the simplest and commonly used unsupervised machine learning algorithms. The objective of K-means is to group similar data points in order to visually discover underlying patterns. K-means clustering starts by picking k centroids (the points around which each cluster will eventually form) and interactively finds points close to each cluster, moving the centroid as needed so that the centroid remains the average point for the cluster.

Hierarchical clustering also creates clusters/categories that have a hierarchical structure. In this case, the top of the hierarchy comprises the most general categories, and the subclusters have more specific categories. Hierarchical clusters begin with a number of clusters and proceed to merge them based on the proximity of cluster members to each other. Which of these two clustering techniques are to be chosen depends on the type of result the data scientist wants for the problem at hand. If you are trying to determine the best ads to display to a user of an online site with a general category such as clothing or furniture, then k-means clustering will be chosen. But if the company wants to have more specific categories as well as general ones (i.e., furniture and kitchen tables), then hierarchical clustering will be useful.

How are unsupervised algorithms evaluated for accuracy? Generally, creating a gold standard corpus is too labor intensive. For clustering techniques, one can measure how close each piece of data is to other pieces of data in the cluster and how far the centroids of each cluster are to other clusters. A number of other very mathematically sophisticated techniques have been explored by data scientists and will not be explained here.

Since all black box algorithms are non-inspectable, must it always be the case that you can’t learn how they work? There is considerable ongoing research to determine scalable and predictable ways to inspect black box algorithms. However, thus far no single approach has emerged as the solution to the interpretability or explainability of models.

There are some important algorithms that function like a combination of supervised and unsupervised models. The most important of these are reinforcement learning and neural networks. Reinforcement learning is a model that receives feedback on its performance to guide it to a good outcome. The system learns based on trial and error. Neural networks are designed to emulate human brains. These networks consist of input nodes, hidden layers, and output nodes. Neural networks are being used in recommendation engines and speech recognition.

Reinforcement learning (RL) has only recently been used in business applications because the way it learns is not so easily adapted to business problems. In RL, the algorithm experiences positive and negative rewards as it makes choices and develops rules inside the model (its choices are called policies). Suppose that the algorithm is playing a simulated game of soccer, where it can kick the ball to another team player, try to kick a goal, or try to steal the ball from an opponent. It chooses an action based on features of the environment (which it chooses for itself) and then takes one of the three actions. If it fails, it gets a negative reward; if it succeeds, it gets a positive reward. It updates the reasons (the policies) for its choices based on the rewards and keeps going. Clearly, game-playing programs benefit from using RL, but other recent business applications have started to appear. For example, suppose our cable television company uses its unsupervised algorithm to decide which types of advertisements to show its customers and now it must bid on ads from several different companies that create ads. An RL model could decide on the basis of the cost of ads, length of ads, and whatever else it determines relevant to which ad to bid on. Then its reward is whether the customer watches the ad (positive reward), turns off the sound (negative reward), or turns off the show (negative reward). Based on those positive and negative rewards, it adjusts its policies for the next ad choices.

Neural network algorithms include so called “deep” neural network algorithms. These algorithms consist of a set of calculations done on “nodes,” each of which is linked to other nodes. In simple neural nets, there are a layer of input nodes, one middle layer of nodes (called the hidden layer), and one layer of output nodes. Figure 4-2 shows connections between the top nodes of the network and all the other nodes. Although a typical network only has three rows of nodes, it is possible to add many more.

This complex network was designed to work something like the neurons in the human brain, though these artificial neural networks are vastly simpler than human neurons and their networks.

The mathematics of neural network algorithms works in the following way. Each node has weights for each input value of that node and a summation function for all the weights. Each summation function has an activation function that determines what output will be provided for that node and its various inputs. Thus, every node in the network has summation and activation functions all making guesses at the correct answer and then adjusting the weights to reduce the error when the node guesses incorrectly.

The neural network is designed to optimize the weights to reduce the error in a correct output. Simple neural networks were, for much of their history, used as supervised learning algorithms, and so the outputs were known for each input. Much of the research in machine learning has explored a wide range of optimization algorithms and activation functions, which will not be discussed in this chapter. In addition, several more specialized types of neural networks have been invented, including feedforward and self-organizing map neural networks, to name just a few. Neural networks have been used for classification problems and clustering (the similarities of some set). Unsupervised simple neural networks have been used to recognize faces, perform speech recognition, make predictions for cancer diagnosis, and for long-term financial forecasting, to name a few examples.

American Express was one of the first companies in the 1990s to use machine learning algorithms to detect fraud in credit card use. American Express evolved their credit and fraud risk management practice from a manual process based on training individuals into a data- and science-based organization.¹ The American Express project was intended to identify patterns of fraud in credit card transactions. The results proved that machine learning models could in fact identify fraud. The value came from American Express’s ability to feed the system with massive amounts of the right data. By detecting patterns that would result in fraud, American Express saved an enormous amount of money. To this day, American Express has continued to heavily invest in machine learning technologies and has industry-leading credit and fraud write-off rates. The success of the project was one of the early indications that machine learning and pattern-based algorithms could become an engine for business transformation. In addition to preventing fraud, American Express was able to offer credit to a broader range of individuals.

Currently, neural network algorithms have been devised to detect credit card fraud in real-time. However, such detection continues to be an active area of work for credit card companies who naturally do not share their algorithms publicly.

In recent years, attention has shifted to deep neural networks and the associated notion of deep learning. Rather than one hidden layer, deep neural networks have many hidden layers, only limited by the compute power to manage the computations. In essence, there is no theoretical way to define the correct number of layers for a given problem. The idea of these many layers is to have each layer train on a given feature and then pass its results to the next layer, which trains on another feature. Deep neural networks are put forward as dealing with highly complex data and finding structure in that data without supervision. Both simple and deep neural network algorithms require substantial data to do their decision making, and for deep neural networks, the amount of data needed is immense. In addition, the amount of computation for large simple neural nets and for any deep neural net can be extensive, so data and computation are major factors in using these kinds of learning algorithms.