Machine learning is most successful when it augments rather than replaces the specialized knowledge of a subject-matter expert. It works with medical doctors at the forefront of the fight to eradicate cancer, assists engineers and programmers with our efforts to create smarter homes and automobiles, and helps social scientists build knowledge of how societies function. Toward these ends, it is employed in countless businesses, scientific laboratories, hospitals, and governmental organizations. Any organization that generates or aggregates data likely employs at least one machine learning algorithm to help make sense of it.

Though it is impossible to list every use case of machine learning, a survey of recent success stories includes several prominent applications:

By the end of this book, you will understand the basic machine learning algorithms that are employed to teach computers to perform these tasks. For now, it suffices to say that no matter what the context is, the machine learning process is the same. Regardless of the task, an algorithm takes data and identifies patterns that form the basis for further action.

Although machine learning is used widely and has tremendous potential, it is important to understand its limits. Machine learning, at this time, is not in any way a substitute for a human brain. It has very little flexibility to extrapolate outside of the strict parameters it learned and knows no common sense. With this in mind, one should be extremely careful to recognize exactly what the algorithm has learned before setting it loose in the real-world settings.

Without a lifetime of past experiences to build upon, computers are also limited in their ability to make simple common sense inferences about logical next steps. Take, for instance, the banner advertisements seen on many web sites. These may be served, based on the patterns learned by data mining the browsing history of millions of users. According to this data, someone who views the websites selling shoes should see advertisements for shoes, and those viewing websites for mattresses should see advertisements for mattresses. The problem is that this becomes a never-ending cycle in which additional shoe or mattress advertisements are served rather than advertisements for shoelaces and shoe polish, or bed sheets and blankets.

Many are familiar with the deficiencies of machine learning's ability to understand or translate language or to recognize speech and handwriting. Perhaps the earliest example of this type of failure is in a 1994 episode of the television show, The Simpsons, which showed a parody of the Apple Newton tablet. For its time, the Newton was known for its state-of-the-art handwriting recognition. Unfortunately for Apple, it would occasionally fail to great effect. The television episode illustrated this through a sequence in which a bully's note to Beat up Martin was misinterpreted by the Newton as Eat up Martha, as depicted in the following screenshots:

Screenshots from "Lisa on Ice" The Simpsons, 20th Century Fox (1994)

Machines' ability to understand language has improved enough since 1994, such that Google, Apple, and Microsoft are all confident enough to offer virtual concierge services operated via voice recognition. Still, even these services routinely struggle to answer relatively simple questions. Even more, online translation services sometimes misinterpret sentences that a toddler would readily understand. The predictive text feature on many devices has also led to a number of humorous autocorrect fail sites that illustrate the computer's ability to understand basic language but completely misunderstand context.

Some of these mistakes are to be expected, for sure. Language is complicated with multiple layers of text and subtext and even human beings, sometimes, understand the context incorrectly. This said, these types of failures in machines illustrate the important fact that machine learning is only as good as the data it learns from. If the context is not directly implicit in the input data, then just like a human, the computer will have to make its best guess.

At its core, machine learning is simply a tool that assists us in making sense of the world's complex data. Like any tool, it can be used for good or evil. Machine learning may lead to problems when it is applied so broadly or callously that humans are treated as lab rats, automata, or mindless consumers. A process that may seem harmless may lead to unintended consequences when automated by an emotionless computer. For this reason, those using machine learning or data mining would be remiss not to consider the ethical implications of the art.

Due to the relative youth of machine learning as a discipline and the speed at which it is progressing, the associated legal issues and social norms are often quite uncertain and constantly in flux. Caution should be exercised while obtaining or analyzing data in order to avoid breaking laws, violating terms of service or data use agreements, and abusing the trust or violating the privacy of customers or the public.

Retailers routinely use machine learning for advertising, targeted promotions, inventory management, or the layout of the items in the store. Many have even equipped checkout lanes with devices that print coupons for promotions based on the customer's buying history. In exchange for a bit of personal data, the customer receives discounts on the specific products he or she wants to buy. At first, this appears relatively harmless. But consider what happens when this practice is taken a little bit further.

One possibly apocryphal tale concerns a large retailer in the U.S. that employed machine learning to identify expectant mothers for coupon mailings. The retailer hoped that if these mothers-to-be received substantial discounts, they would become loyal customers, who would later purchase profitable items like diapers, baby formula, and toys.

Equipped with machine learning methods, the retailer identified items in the customer purchase history that could be used to predict with a high degree of certainty, not only whether a woman was pregnant, but also the approximate timing for when the baby was due.

After the retailer used this data for a promotional mailing, an angry man contacted the chain and demanded to know why his teenage daughter received coupons for maternity items. He was furious that the retailer seemed to be encouraging teenage pregnancy! As the story goes, when the retail chain's manager called to offer an apology, it was the father that ultimately apologized because, after confronting his daughter, he discovered that she was indeed pregnant!

Whether completely true or not, the lesson learned from the preceding tale is that common sense should be applied before blindly applying the results of a machine learning analysis. This is particularly true in cases where sensitive information such as health data is concerned. With a bit more care, the retailer could have foreseen this scenario, and used greater discretion while choosing how to reveal the pattern its machine learning analysis had discovered.

Certain jurisdictions may prevent you from using racial, ethnic, religious, or other protected class data for business reasons. Keep in mind that excluding this data from your analysis may not be enough, because machine learning algorithms might inadvertently learn this information independently. For instance, if a certain segment of people generally live in a certain region, buy a certain product, or otherwise behave in a way that uniquely identifies them as a group, some machine learning algorithms can infer the protected information from these other factors. In such cases, you may need to fully "de-identify" these people by excluding any potentially identifying data in addition to the protected information.

Apart from the legal consequences, using data inappropriately may hurt the bottom line. Customers may feel uncomfortable or become spooked if the aspects of their lives they consider private are made public. In recent years, several high-profile web applications have experienced a mass exodus of users who felt exploited when the applications' terms of service agreements changed, and their data was used for purposes beyond what the users had originally agreed upon. The fact that privacy expectations differ by context, age cohort, and locale adds complexity in deciding the appropriate use of personal data. It would be wise to consider the cultural implications of your work before you begin your project.

All learning must begin with data. Humans and computers alike utilize data storage as a foundation for more advanced reasoning. In a human being, this consists of a brain that uses electrochemical signals in a network of biological cells to store and process observations for short- and long-term future recall. Computers have similar capabilities of short- and long-term recall using hard disk drives, flash memory, and random access memory (RAM) in combination with a central processing unit (CPU).

It may seem obvious to say so, but the ability to store and retrieve data alone is not sufficient for learning. Without a higher level of understanding, knowledge is limited exclusively to recall, meaning exclusively what is seen before and nothing else. The data is merely ones and zeros on a disk. They are stored memories with no broader meaning.

To better understand the nuances of this idea, it may help to think about the last time you studied for a difficult test, perhaps for a university final exam or a career certification. Did you wish for an eidetic (photographic) memory? If so, you may be disappointed to learn that perfect recall is unlikely to be of much assistance. Even if you could memorize material perfectly, your rote learning is of no use, unless you know in advance the exact questions and answers that will appear in the exam. Otherwise, you would be stuck in an attempt to memorize answers to every question that could conceivably be asked. Obviously, this is an unsustainable strategy.

Instead, a better approach is to spend time selectively, memorizing a small set of representative ideas while developing strategies on how the ideas relate and how to use the stored information. In this way, large ideas can be understood without needing to memorize them by rote.

This work of assigning meaning to stored data occurs during the abstraction process, in which raw data comes to have a more abstract meaning. This type of connection, say between an object and its representation, is exemplified by the famous René Magritte painting The Treachery of Images:

Source: http://collections.lacma.org/node/239578

The painting depicts a tobacco pipe with the caption Ceci n'est pas une pipe ("this is not a pipe"). The point Magritte was illustrating is that a representation of a pipe is not truly a pipe. Yet, in spite of the fact that the pipe is not real, anybody viewing the painting easily recognizes it as a pipe. This suggests that the observer's mind is able to connect the picture of a pipe to the idea of a pipe, to a memory of a physical pipe that could be held in the hand. Abstracted connections like these are the basis of knowledge representation, the formation of logical structures that assist in turning raw sensory information into a meaningful insight.

During a machine's process of knowledge representation, the computer summarizes stored raw data using a model, an explicit description of the patterns within the data. Just like Magritte's pipe, the model representation takes on a life beyond the raw data. It represents an idea greater than the sum of its parts.

There are many different types of models. You may be already familiar with some. Examples include:

The choice of model is typically not left up to the machine. Instead, the learning task and data on hand inform model selection. Later in this chapter, we will discuss methods to choose the type of model in more detail.

The process of fitting a model to a dataset is known as training. When the model has been trained, the data is transformed into an abstract form that summarizes the original information.

It is important to note that a learned model does not itself provide new data, yet it does result in new knowledge. How can this be? The answer is that imposing an assumed structure on the underlying data gives insight into the unseen by supposing a concept about how data elements are related. Take for instance the discovery of gravity. By fitting equations to observational data, Sir Isaac Newton inferred the concept of gravity. But the force we now know as gravity was always present. It simply wasn't recognized until Newton recognized it as an abstract concept that relates some data to others—specifically, by becoming the g term in a model that explains observations of falling objects.

Most models may not result in the development of theories that shake up scientific thought for centuries. Still, your model might result in the discovery of previously unseen relationships among data. A model trained on genomic data might find several genes that, when combined, are responsible for the onset of diabetes; banks might discover a seemingly innocuous type of transaction that systematically appears prior to fraudulent activity; and psychologists might identify a combination of personality characteristics indicating a new disorder. These underlying patterns were always present, but by simply presenting information in a different format, a new idea is conceptualized.

The learning process is not complete until the learner is able to use its abstracted knowledge for future action. However, among the countless underlying patterns that might be identified during the abstraction process and the myriad ways to model these patterns, some will be more useful than others. Unless the production of abstractions is limited, the learner will be unable to proceed. It would be stuck where it started—with a large pool of information, but no actionable insight.

The term generalization describes the process of turning abstracted knowledge into a form that can be utilized for future action, on tasks that are similar, but not identical, to those it has seen before. Generalization is a somewhat vague process that is a bit difficult to describe. Traditionally, it has been imagined as a search through the entire set of models (that is, theories or inferences) that could be abstracted during training. In other words, if you can imagine a hypothetical set containing every possible theory that could be established from the data, generalization involves the reduction of this set into a manageable number of important findings.

In generalization, the learner is tasked with limiting the patterns it discovers to only those that will be most relevant to its future tasks. Generally, it is not feasible to reduce the number of patterns by examining them one-by-one and ranking them by future utility. Instead, machine learning algorithms generally employ shortcuts that reduce the search space more quickly. Toward this end, the algorithm will employ heuristics, which are educated guesses about where to find the most useful inferences.

Heuristics are routinely used by human beings to quickly generalize experience to new scenarios. If you have ever utilized your gut instinct to make a snap decision prior to fully evaluating your circumstances, you were intuitively using mental heuristics.

The incredible human ability to make quick decisions often relies not on computer-like logic, but rather on heuristics guided by emotions. Sometimes, this can result in illogical conclusions. For example, more people express fear of airline travel versus automobile travel, despite automobiles being statistically more dangerous. This can be explained by the availability heuristic, which is the tendency of people to estimate the likelihood of an event by how easily its examples can be recalled. Accidents involving air travel are highly publicized. Being traumatic events, they are likely to be recalled very easily, whereas car accidents barely warrant a mention in the newspaper.

The folly of misapplied heuristics is not limited to human beings. The heuristics employed by machine learning algorithms also sometimes result in erroneous conclusions. The algorithm is said to have a bias if the conclusions are systematically erroneous, or wrong in a predictable manner.

For example, suppose that a machine learning algorithm learned to identify faces by finding two dark circles representing eyes, positioned above a straight line indicating a mouth. The algorithm might then have trouble with, or be biased against, faces that do not conform to its model. Faces with glasses, turned at an angle, looking sideways, or with various skin tones might not be detected by the algorithm. Similarly, it could be biased toward faces with certain skin tones, face shapes, or other characteristics that do not conform to its understanding of the world.

In modern usage, the word bias has come to carry quite negative connotations. Various forms of media frequently claim to be free from bias, and claim to report the facts objectively, untainted by emotion. Still, consider for a moment the possibility that a little bias might be useful. Without a bit of arbitrariness, might it be a bit difficult to decide among several competing choices, each with distinct strengths and weaknesses? Indeed, some recent studies in the field of psychology have suggested that individuals born with damage to portions of the brain responsible for emotion are ineffectual in decision making, and might spend hours debating simple decisions such as what color shirt to wear or where to eat lunch. Paradoxically, bias is what blinds us from some information while also allowing us to utilize other information for action. It is how machine learning algorithms choose among the countless ways to understand a set of data.

Bias is a necessary evil associated with the abstraction and generalization processes inherent in any learning task. In order to drive action in the face of limitless possibility, each learner must be biased in a particular way. Consequently, each learner has its weaknesses and there is no single learning algorithm to rule them all. Therefore, the final step in the generalization process is to evaluate or measure the learner's success in spite of its biases and use this information to inform additional training if needed.

Generally, evaluation occurs after a model has been trained on an initial training dataset. Then, the model is evaluated on a new test dataset in order to judge how well its characterization of the training data generalizes to new, unseen data. It's worth noting that it is exceedingly rare for a model to perfectly generalize to every unforeseen case.

In parts, models fail to perfectly generalize due to the problem of noise, a term that describes unexplained or unexplainable variations in data. Noisy data is caused by seemingly random events, such as:

Trying to model noise is the basis of a problem called overfitting. Because most noisy data is unexplainable by definition, attempting to explain the noise will result in erroneous conclusions that do not generalize well to new cases. Efforts to explain the noise will also typically result in more complex models that will miss the true pattern that the learner tries to identify. A model that seems to perform well during training, but does poorly during evaluation, is said to be overfitted to the training dataset, as it does not generalize well to the test dataset.

Solutions to the problem of overfitting are specific to particular machine learning approaches. For now, the important point is to be aware of the issue. How well the models are able to handle noisy data is an important source of distinction among them.

The practice of machine learning involves matching the characteristics of input data to the biases of the available approaches. Thus, before applying machine learning to real-world problems, it is important to understand the terminology that distinguishes among input datasets.

The phrase unit of observation is used to describe the smallest entity with measured properties of interest for a study. Commonly, the unit of observation is in the form of persons, objects or things, transactions, time points, geographic regions, or measurements. Sometimes, units of observation are combined to form units such as person-years, which denote cases where the same person is tracked over multiple years; each person-year comprises of a person's data for one year.

Datasets that store the units of observation and their properties can be imagined as collections of data consisting of:

It is easiest to understand features and examples through real-world cases. To build a learning algorithm to identify spam e-mail, the unit of observation could be e-mail messages, the examples would be specific messages, and the features might consist of the words used in the messages. For a cancer detection algorithm, the unit of observation could be patients, the examples might include a random sample of cancer patients, and the features may be the genomic markers from biopsied cells as well as the characteristics of patient such as weight, height, or blood pressure.

While examples and features do not have to be collected in any specific form, they are commonly gathered in matrix format, which means that each example has exactly the same features.

The following spreadsheet shows a dataset in matrix format. In matrix data, each row in the spreadsheet is an example and each column is a feature. Here, the rows indicate examples of automobiles, while the columns record various each automobile's features, such as price, mileage, color, and transmission type. Matrix format data is by far the most common form used in machine learning. Though, as you will see in the later chapters, other forms are used occasionally in specialized cases:

Features also come in various forms. If a feature represents a characteristic measured in numbers, it is unsurprisingly called numeric. Alternatively, if a feature is an attribute that consists of a set of categories, the feature is called categorical or nominal. A special case of categorical variables is called ordinal, which designates a nominal variable with categories falling in an ordered list. Some examples of ordinal variables include clothing sizes such as small, medium, and large; or a measurement of customer satisfaction on a scale from "not at all happy" to "very happy." It is important to consider what the features represent, as the type and number of features in your dataset will assist in determining an appropriate machine learning algorithm for your task.

Machine learning algorithms are divided into categories according to their purpose. Understanding the categories of learning algorithms is an essential first step towards using data to drive the desired action.

A predictive model is used for tasks that involve, as the name implies, the prediction of one value using other values in the dataset. The learning algorithm attempts to discover and model the relationship between the target feature (the feature being predicted) and the other features. Despite the common use of the word "prediction" to imply forecasting, predictive models need not necessarily foresee events in the future. For instance, a predictive model could be used to predict past events, such as the date of a baby's conception using the mother's present-day hormone levels. Predictive models can also be used in real time to control traffic lights during rush hours.

Because predictive models are given clear instruction on what they need to learn and how they are intended to learn it, the process of training a predictive model is known as supervised learning. The supervision does not refer to human involvement, but rather to the fact that the target values provide a way for the learner to know how well it has learned the desired task. Stated more formally, given a set of data, a supervised learning algorithm attempts to optimize a function (the model) to find the combination of feature values that result in the target output.

The often used supervised machine learning task of predicting which category an example belongs to is known as classification. It is easy to think of potential uses for a classifier. For instance, you could predict whether:

In classification, the target feature to be predicted is a categorical feature known as the class, and is divided into categories called levels. A class can have two or more levels, and the levels may or may not be ordinal. Because classification is so widely used in machine learning, there are many types of classification algorithms, with strengths and weaknesses suited for different types of input data. We will see examples of these later in this chapter and throughout this book.

Supervised learners can also be used to predict numeric data such as income, laboratory values, test scores, or counts of items. To predict such numeric values, a common form of numeric prediction fits linear regression models to the input data. Although regression models are not the only type of numeric models, they are, by far, the most widely used. Regression methods are widely used for forecasting, as they quantify in exact terms the association between inputs and the target, including both, the magnitude and uncertainty of the relationship.

A descriptive model is used for tasks that would benefit from the insight gained from summarizing data in new and interesting ways. As opposed to predictive models that predict a target of interest, in a descriptive model, no single feature is more important than any other. In fact, because there is no target to learn, the process of training a descriptive model is called unsupervised learning. Although it can be more difficult to think of applications for descriptive models—after all, what good is a learner that isn't learning anything in particular—they are used quite regularly for data mining.

For example, the descriptive modeling task called pattern discovery is used to identify useful associations within data. Pattern discovery is often used for market basket analysis on retailers' transactional purchase data. Here, the goal is to identify items that are frequently purchased together, such that the learned information can be used to refine marketing tactics. For instance, if a retailer learns that swimming trunks are commonly purchased at the same time as sunglasses, the retailer might reposition the items more closely in the store or run a promotion to "up-sell" customers on associated items.

The descriptive modeling task of dividing a dataset into homogeneous groups is called clustering. This is sometimes used for segmentation analysis that identifies groups of individuals with similar behavior or demographic information, so that advertising campaigns could be tailored for particular audiences. Although the machine is capable of identifying the clusters, human intervention is required to interpret them. For example, given five different clusters of shoppers at a grocery store, the marketing team will need to understand the differences among the groups in order to create a promotion that best suits each group.

Lastly, a class of machine learning algorithms known as meta-learners is not tied to a specific learning task, but is rather focused on learning how to learn more effectively. A meta-learning algorithm uses the result of some learnings to inform additional learning. This can be beneficial for very challenging problems or when a predictive algorithm's performance needs to be as accurate as possible.

The following table lists the general types of machine learning algorithms covered in this book. Although this covers only a fraction of the entire set of machine learning algorithms, learning these methods will provide a sufficient foundation to make sense of any other method you may encounter in the future.

To begin applying machine learning to a real-world project, you will need to determine which of the four learning tasks your project represents: classification, numeric prediction, pattern detection, or clustering. The task will drive the choice of algorithm. For instance, if you are undertaking pattern detection, you are likely to employ association rules. Similarly, a clustering problem will likely utilize the k-means algorithm, and numeric prediction will utilize regression analysis or regression trees.

For classification, more thought is needed to match a learning problem to an appropriate classifier. In these cases, it is helpful to consider various distinctions among algorithms—distinctions that will only be apparent by studying each of the classifiers in depth. For instance, within classification problems, decision trees result in models that are readily understood, while the models of neural networks are notoriously difficult to interpret. If you were designing a credit-scoring model, this could be an important distinction because law often requires that the applicant must be notified about the reasons he or she was rejected for the loan. Even if the neural network is better at predicting loan defaults, if its predictions cannot be explained, then it is useless for this application.

To assist with the algorithm selection, in every chapter, the key strengths and weaknesses of each learning algorithm are listed. Although you will sometimes find that these characteristics exclude certain models from consideration, in many cases, the choice of algorithm is arbitrary. When this is true, feel free to use whichever algorithm you are most comfortable with. Other times, when predictive accuracy is primary, you may need to test several algorithms and choose the one that fits the best, or use a meta-learning algorithm that combines several different learners to utilize the strengths of each.

Despite the vast set of available R add-ons, the package format makes installation and use a virtually effortless process. To demonstrate the use of packages, we will install and load the RWeka package, which was developed by Kurt Hornik, Christian Buchta, and Achim Zeileis (see Open-Source Machine Learning: R Meets Weka in Computational Statistics 24: 225-232 for more information). The RWeka package provides a collection of functions that give R access to the machine learning algorithms in the Java-based Weka software package by Ian H. Witten and Eibe Frank. More information on Weka is available at http://www.cs.waikato.ac.nz/~ml/weka/.

The most direct way to install a package is via the install.packages() function. To install the RWeka package, at the R command prompt, simply type:

> install.packages("RWeka")

R will then connect to CRAN and download the package in the correct format for your OS. Some packages such as RWeka require additional packages to be installed before they can be used (these are called dependencies). By default, the installer will automatically download and install any dependencies.

The default installation options are appropriate for most systems. However, in some cases, you may want to install a package to another location. For example, if you do not have root or administrator privileges on your system, you may need to specify an alternative installation path. This can be accomplished using the lib option, as follows:

> install.packages("RWeka", lib="/path/to/library")

The installation function also provides additional options for installation from a local file, installation from source, or using experimental versions. You can read about these options in the help file, by using the following command:

> ?install.packages

More generally, the question mark operator can be used to obtain help on any R function. Simply type ? before the name of the function.

In order to conserve memory, R does not load every installed package by default. Instead, packages are loaded by users as they are needed, using the library() function.

To load the RWeka package we installed previously, you can type the following:

> library(RWeka)

Aside from RWeka, there are several other R packages that will be used in the later chapters. Installation instructions will be provided as additional packages are used.

To unload an R package, use the detach() function. For example, to unload the RWeka package shown previously use the following command:

> detach("package:RWeka", unload = TRUE)

This will free up any resources used by the package.