Nowadays, everyone talks about machine learning and data science. So, what exactly is machine learning, anyway? How does it relate to data science? These two terms are commonly confused, as they often employ the same methods and overlap significantly. Therefore, let's first clarify what they are. Josh Wills tweeted this:

More specifically, data science encompasses the entire process of obtaining knowledge by integrating methods from statistics, computer science, and other fields to gain insight from data. In practice, data science encompasses an iterative process of data harvesting, cleaning, analysis, visualization, and deployment.

Machine learning, on the other hand, is mainly concerned with generic algorithms and techniques that are used in analysis and modelling phases of the data science process.

Among the different machine learning approaches, there are three main ways of learning, as shown in the following list:

• Supervised learning
• Unsupervised learning
• Reinforcement learning

Given a set of example inputs X, and their outcomes Y, supervised learning aims to learn a general mapping function f, which transforms inputs into outputs, as f: (X,Y).

An example of supervised learning is credit card fraud detection, where the learning algorithm is presented with credit card transactions (matrix X) marked as normal or suspicious (vector Y). The learning algorithm produces a decision model that marks unseen transactions as normal or suspicious (this is the f function).

In contrast, unsupervised learning algorithms do not assume given outcome labels, as they focus on learning the structure of the data, such as grouping similar inputs into clusters. Unsupervised learning can, therefore, discover hidden patterns in the data. An example of unsupervised learning is an item-based recommendation system, where the learning algorithm discovers similar items bought together; for example, people who bought book A also bought book B.

Reinforcement learning addresses the learning process from a completely different angle. It assumes that an agent, which can be a robot, bot, or computer program, interacts with a dynamic environment to achieve a specific goal. The environment is described with a set of states and the agent can take different actions to move from one state to another. Some states are marked as goal states, and if the agent achieves this state, it receives a large reward. In other states, the reward is smaller, non-existent, or even negative. The goal of reinforcement learning is to find an optimal policy or a mapping function that specifies the action to take in each of the states, without a teacher explicitly telling whether this leads to the goal state or not. An example of reinforcement learning would be a program for driving a vehicle, where the states correspond to the driving conditions, for example, current speed, road segment information, surrounding traffic, speed limits, and obstacles on the road; and the actions could be driving maneuvers, such as turn left or right, stop, accelerate, and continue. The learning algorithm produces a policy that specifies the action that is to be taken in specific configurations of driving conditions.

In this book, we will focus on supervised and unsupervised learning only, as they share many concepts. If reinforcement learning sparked your interest, a good book to start with is Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew Barto, MIT Press (2018).

This book's emphasis is on applied machine learning. We want to provide you with the practical skills needed to get learning algorithms to work in different settings. Instead of math and theory in machine learning, we will spend more time on the practical, hands-on skills (and dirty tricks) to get this stuff to work well on an application. We will focus on supervised and unsupervised machine learning and learn the essential steps in data science to build the applied machine learning workflow.

A typical workflow in applied machine learning applications consists of answering a series of questions that can be summarized in the following steps:

1. Data and problem definition: The first step is to ask interesting questions, such as: What is the problem you are trying solve? Why is it important? Which format of result answers your question? Is this a simple yes/no answer? Do you need to pick one of the available questions?
2. Data collection: Once you have a problem to tackle, you will need the data. Ask yourself what kind of data will help you answer the question. Can you get the data from the available sources? Will you have to combine multiple sources? Do you have to generate the data? Are there any sampling biases? How much data will be required?
3. Data preprocessing: The first data preprocessing task is data cleaning. Some of the examples include filling missing values, smoothing noisy data, removing outliers, and resolving consistencies. This is usually followed by integration of multiple data sources and data transformation to a specific range (normalization), to value bins (discretized intervals), and to reduce the number of dimensions.
4. Data analysis and modelling: Data analysis and modelling includes unsupervised and supervised machine learning, statistical inference, and prediction. A wide variety of machine learning algorithms are available, including k-nearest neighbors, Naive Bayes classifier, decision trees, Support Vector Machines (SVMs), logistic regression, k-means, and so on. The method to be deployed depends on the problem definition, as discussed in the first step, and the type of collected data. The final product of this step is a model inferred from the data.
5. Evaluation: The last step is devoted to model assessment. The main issue that the models built with machine learning face is how well they model the underlying data; for example, if a model is too specific or it overfits to the data used for training, it is quite possible that it will not perform well on new data. The model can be too generic, meaning that it underfits the training data. For example, when asked how the weather is in California, it always answers sunny, which is indeed correct most of the time. However, such a model is not really useful for making valid predictions. The goal of this step is to correctly evaluate the model and make sure it will work on new data as well. Evaluation methods include separate test and train sets, cross-validation, and leave-one-out cross-validation.

We will take a closer look at each of the steps in the following sections. We will try to understand the type of questions we must answer during the applied machine learning workflow, and look at the accompanying concepts of data analysis and evaluation.

When presented with a problem definition, we need to ask questions that will help in understanding the objective and target information from the data. We could ask very common questions, such as: what is the expected finding once the data is explored? What kind of information can be extracted after data exploration? Or, what kind of format is required so the question can be answered? Asking the right question will give a clearer understanding of how to proceed further. Data is simply a collection of measurements in the form of numbers, words, observations, descriptions of things, images, and more.

The most common way to represent data is using a set of attribute-value pairs. Consider the following example:

Bob = { 
height: 185cm, 
eye color: blue, 
hobbies: climbing, sky diving 
} 

For example, Bob has attributes named height, eye color, and hobbies with the values 185cm, blue, climbing, and sky diving respectively.

A set of data can be presented simply as a table, where columns correspond to attributes or features and rows correspond to particular data examples or instances. In supervised machine learning, the attribute whose value we want to predict the outcome, Y, from the values of the other attributes, X, is denoted as the class or target variable, as shown in the following table:

The first thing we notice is how much the attribute values vary. For instance, height is a number, eye color is text, and hobbies are a list. To gain a better understanding of the value types, let's take a closer look at the different types of data or measurement scales. Stanley Smith Stevens (1946) defined the following four scales of measurement with increasingly expressive properties:

• Nominal data consists of data that is mutually exclusive, but not ordered. Examples include eye color, marital status, type of car owned, and so on.
• Ordinal data correspond to categories where order matters, but not the difference between the values, such as pain level, student letter grades, service quality ratings, IMDb movie ratings, and so on.
• Interval data consists of data where the difference between two values is meaningful, but there is no concept of zero, for instance, standardized exam scores, temperature in Fahrenheit, and so on.
• Ratio data has all of the properties of an interval variable and also a clear definition of zero; when the variable is equal to zero, this variable would be missing. Variables such as height, age, stock prices, and weekly food spending are ratio variables.

Why should we care about measurement scales? Well, machine learning depends heavily on the statistical properties of the data; hence, we should be aware of the limitations that each data type possesses. Some machine learning algorithms can only be applied to a subset of measurement scales.

The following table summarizes the main operations and statistics properties for each of the measurement types:

Furthermore, nominal and ordinal data correspond to discrete values, while interval and ratio data can correspond to continuous values as well. In supervised learning, the measurement scale of the attribute values that we want to predict dictates the kind of machine algorithm that can be used. For instance, predicting discrete values from a limited list is called classification and can be achieved using decision trees, while predicting continuous values is called regression, which can be achieved using model trees.

Once questions are asked in the right direction, the target of data exploration is clear. So, the next step is to see where the data comes from. Data collected can be much unorganized and in very diverse formats, which may involve reading from a database, internet, file system, or other documents. Most of the tools for machine learning require data to be presented in a specific format in order to generate the proper result. We have two choices: observe the data from existing sources or generate the data via surveys, simulations, and experiments. Let's take a closer look at both approaches.

Data can be found or observed in many places. An obvious data source is the internet. With an increase in social media usage, and with mobile phones penetrating deeper as mobile data plans become cheaper or even offer unlimited data, there has been an exponential rise in data consumed by users.

Now, online streaming platforms have emerged—the following diagram shows that the hours spent on consuming video data is also growing rapidly:

To get data from the internet, there are multiple options, as shown in the following list:

• Bulk downloads from websites such as Wikipedia, IMDb, and the Million Song Dataset (which can be found here: https://labrosa.ee.columbia.edu/millionsong/).
• Accessing the data through APIs (such as Google, Twitter, Facebook, and YouTube).
• It is okay to scrape public, non-sensitive, and anonymized data. Be sure to check the terms and conditions and to fully reference the information.

The main drawbacks of the data collected is that it takes time and space to accumulate the data, and it covers only what happened; for instance, intentions and internal and external motivations are not collected. Finally, such data might be noisy, incomplete, inconsistent, and may even change over time.

Another option is to collect measurements from sensors such as inertial and location sensors in mobile devices, environmental sensors, and software agents monitoring key performance indicators.

An alternative approach is to generate the data by you, for example, with a survey. In survey design, we have to pay attention to data sampling; that is, who the respondents are that are answering the survey. We only get data from the respondents who are accessible and willing to respond. Also, respondents can provide answers that are in line with their self-image and researcher's expectations.

Alternatively, the data can be collected with simulations, where a domain expert specifies the behavior model of users at a micro level. For instance, crowd simulation requires specifying how different types of users will behave in a crowd. Some of the examples could be following the crowd, looking for an escape, and so on. The simulation can then be run under different conditions to see what happens (Tsai et al., 2011). Simulations are appropriate for studying macro phenomena and emergent behavior; however, they are typically hard to validate empirically.

Furthermore, you can design experiments to thoroughly cover all of the possible outcomes, where you keep all of the variables constant and only manipulate one variable at a time. This is the most costly approach, but usually provides the best quality.

Data collection may involve many traps. To demonstrate one, let me share a story. There is supposed to be a global, unwritten rule for sending regular mail between students for free. If you write student to student in the place where the stamp should be, the mail is delivered to the recipient for free. Now, suppose Jacob sends a set of postcards to Emma, and given that Emma indeed receives some of the postcards, she concludes that all of the postcards are delivered and that the rule indeed holds true. Emma reasons that, as she received the postcards, all of the postcards are delivered. However, she does not know of the postcards that were sent by Jacob, but were undelivered; hence, she is unable to account for this in her inference. What Emma experienced is survivorship bias; that is, she drew the conclusion based on the data that survived. For your information, postcards that are sent with a student to student stamp get a circled black letter T stamp on them, which mean postage is due and the receiver should pay it, including a small fine. However, mail services often have higher costs on applying such fees and hence do not do it. (Magalhães, 2010).

Another example is a study that found that the profession with the lowest average age of death was student. Being a student does not cause you to die at an early age; rather, being a student means you are young. This is why the average is so low. (Gelman and Nolan, 2002).

Furthermore, a study that found that only 1.5% of drivers in accidents reported they were using a cell phone, whereas 10.9% reported another occupant in the car distracted them. Can we conclude that using a cell phone is safer than speaking with another occupant? (Uts, 2003) To answer this question, we need to know the prevalence of the cell phone use. It is likely that a higher number of people talked to another occupant in the car while driving than talked on a cell phone during the period when the data was collected.

The goal of data preprocessing tasks is to prepare the data for a machine learning algorithm in the best possible way, as not all algorithms are capable of addressing issues with missing data, extra attributes, or denormalized values.

Data cleaning, also known as data cleansing or data scrubbing, is a process consisting of the following steps:

1. Identifying inaccurate, incomplete, irrelevant, or corrupted data to remove it from further processing
2. Parsing data, extracting information of interest, or validating whether a string of data is in an acceptable format
3. Transforming data into a common encoding format, for example, UTF-8 or int32, time scale, or a normalized range
4. Transforming data into a common data schema; for instance, if we collect temperature measurements from different types of sensors, we might want them to have the same structure

Machine learning algorithms generally do not work well with missing values. Rare exceptions include decision trees, Naive Bayes classifier, and some rule-based learners. It is very important to understand why a value is missing. It can be missing due to many reasons, such as random error, systematic error, and sensor noise. Once we identify the reason, there are multiple ways to deal with the missing values, as shown in the following list:

• Remove the instance: If there is enough data, and only a couple of non-relevant instances have some missing values, then it is safe to remove these instances.
• Remove the attribute: Removing an attribute makes sense when most of the values are missing, values are constant, or an attribute is strongly correlated with another attribute.
• Assign a special value (N/A): Sometimes a value is missing due to valid reasons, such as the value is out of scope, the discrete attribute value is not defined, or it is not possible to obtain or measure the value. For example, if a person never rates a movie, their rating on this movie is nonexistent.
• Take the average attribute value: If we have a limited number of instances, we might not be able to afford removing instances or attributes. In that case, we can estimate the missing values by assigning the average attribute value.
• Predict the value from other attributes: Predict the value from previous entries if the attribute possesses time dependencies.

As we have seen, the value can be missing for many reasons, and hence, it is important to understand why the value is missing, absent, or corrupted.

Outliers in data are values that are unlike any other values in the series and affect all learning methods to various degrees. These can be extreme values, which could be detected with confidence intervals and removed by using a threshold. The best approach is to visualize the data and inspect the visualization to detect irregularities. An example is shown in the following diagram. Visualization applies to low-dimensional data only:

Data transformation techniques tame the dataset to a format that a machine learning algorithm expects as input and may even help the algorithm to learn faster and achieve better performance. It is also known as data munging or data wrangling. Standardization, for instance, assumes that data follows Gaussian distribution and transforms the values in such a way that the mean value is 0 and the deviation is 1, as follows:

Normalization, on the other hand, scales the values of attributes to a small, specified range, usually between 0 and 1:

Many machine learning toolboxes automatically normalize and standardize the data for you.

The last transformation technique is discretization, which divides the range of a continuous attribute into intervals. Why should we care? Some algorithms, such as decision trees and Naive Bayes prefer discrete attributes. The most common ways to select the intervals are as follows:

• Equal width: The interval of continuous variables is divided into k equal width intervals
• Equal frequency: Supposing there are N instances, each of the k intervals contains approximately N or k instances
• Min entropy: This approach recursively splits the intervals until the entropy, which measures disorder, decreases more than the entropy increase, introduced by the interval split (Fayyad and Irani, 1993)

The first two methods require us to specify the number of intervals, while the last method sets the number of intervals automatically; however, it requires the class variable, which means it won't work for unsupervised machine learning tasks.

Data reduction deals with abundant attributes and instances. The number of attributes corresponds to the number of dimensions in our dataset. Dimensions with low prediction power contribute very little to the overall model, and cause a lot of harm. For instance, an attribute with random values can introduce some random patterns that will be picked up by a machine learning algorithm. It may happen that data contains a large number of missing values, wherein we have to find the reason for missing values in large numbers, and on that basis, it may fill it with some alternate value or impute or remove the attribute altogether. If 40% or more values are missing, then it may be advisable to remove such attributes, as this will impact the model performance.

The other factor is variance, where the constant variable may have low variance, which means the data is very close to each other or there is not very much variation in the data.

To deal with this problem, the first set of techniques removes such attributes and selects the most promising ones. This process is known as feature selection, or attributes selection, and includes methods such as ReliefF, information gain, and the Gini index. These methods are mainly focused on discrete attributes.

Another set of tools, focused on continuous attributes, transforms the dataset from the original dimensions into a lower-dimensional space. For example, if we have a set of points in three-dimensional space, we can make a projection into a two-dimensional space. Some information is lost, but in a situation where the third dimension is irrelevant, we don't lose much, as the data structure and relationships are almost perfectly preserved. This can be performed by the following methods:

• Singular value decomposition (SVD)
• Principal component analysis (PCA)
• Backward/forward feature elimination
• Factor analysis
• Linear discriminant analysis (LDA)
• Neural network autoencoders

The second problem in data reduction is related to too many instances; for example, they can be duplicates or come from a very frequent data stream. The main idea is to select a subset of instances in such a way that distribution of the selected data still resembles the original data distribution, and more importantly, the observed process. Techniques to reduce the number of instances involve random data sampling, stratification, and others. Once the data is prepared, we can start with the data analysis and modeling.

Unsupervised learning is about analyzing the data and discovering hidden structures in unlabeled data. As no notion of the right labels is given, there is also no error measure to evaluate a learned model; however, unsupervised learning is an extremely powerful tool. Have you ever wondered how Amazon can predict what books you'll like? Or how Netflix knows what you want to watch before you do? The answer can be found in unsupervised learning. We will look at a similar example of unsupervised learning in the following section.

Many problems can be formulated as finding similar sets of elements, for example, customers who purchased similar products, web pages with similar content, images with similar objects, users who visited similar websites, and so on.

Two items are considered similar if they are a small distance apart. The main questions are how each item is represented and how the distance between the items is defined. There are two main classes of distance measures:

• Euclidean distances
• Non-Euclidean distances

In Euclidean space, with the n dimension, the distance between two elements is based on the locations of the elements in such a space, which is expressed as p-norm distance. Two commonly used distance measures are L2- and L1-norm distances.

L2-norm, also known as Euclidean distance, is the most frequently applied distance measure that measures how far apart two items in a two-dimensional space are. It is calculated as follows:

L1-norm, also known as Manhattan distance, city block distance, and taxicab norm, simply sums the absolute differences in each dimension, as follows:

A non-Euclidean distance is based on the properties of the elements, but not on their location in space. Some well known distances are Jaccard distance, cosine distance, edit distance, and Hamming distance.

Jaccard distance is used to compute the distance between two sets. First, we compute the Jaccard similarity of two sets as the size of their intersection divided by the size of their union, as follows:

The Jaccard distance is then defined as per the following formula:

Cosine distance between two vectors focuses on the orientation and not magnitude, therefore, two vectors with the same orientation have a cosine similarity of 1, while two perpendicular vectors have a cosine similarity of 0. Supposing that we have two multidimensional points, think of a point as a vector from origin (0, 0, ..., 0) to its location. Two vectors make an angle, whose cosine distance is a normalized dot-product of the vectors, as follows:

Cosine distance is commonly used in a high-dimensional feature space; for instance, in text mining, where a text document represents an instance, features that correspond to different words, and their values correspond to the number of times the word appears in the document. By computing cosine similarity, we can measure how likely two documents match in describing similar content.

Edit distance makes sense when we compare two strings. The distance between the a=a1,a2,a3,...an and b=b1,b2,b3,...bn strings is the smallest number of the insert/delete operation of single characters required to convert the string from a to b, for example, a = abcd and b = abbd. To convert a into b, we have to delete the second b and insert c in its place. No smallest number of operations would convert a into b, hence the distance is d(a, b) =2.

Hamming distance compares two vectors of the same size and counts the number of dimensions in which they differ. In other words, it measures the number of substitutions required to convert one vector into another.

There are many distance measures focusing on various properties, for instance, correlation measures the linear relationship between two elements; Mahalanobis distance measures the distance between a point and distribution of other points and SimRank, which is based on graph theory, measures similarity of the structure in which elements occur, and so on. As you can already imagine, selecting and designing the right similarity measure for your problem is more than half of the battle. An impressive overview and evaluation of similarity measures is collected in Chapter 2, Similarity and Dissimilarity Measures, in the book Image Registration: Principles, Tools and Methods by A. A. Goshtasby, Springer Science and Business Media (2012).

The curse of dimensionality refers to a situation where we have a large number of features, often hundreds or thousands, which lead to an extremely large space with sparse data and, consequently, to distance anomalies. For instance, in high dimensions, almost all pairs of points are equally distant from each other; in fact, almost all of the pairs have distance close to the average distance. Another manifestation of the curse is that any two vectors are almost orthogonal, which means all of the angles are close to 90 degrees. This practically makes any distance measurement useless.

A cure for the curse of dimensionality might be found in one of the data reduction techniques, where we want to reduce the number of features; for instance, we can run a feature selection algorithm, such as ReliefF, or a feature extraction or reduction algorithm, such as PCA.

Clustering is a technique for grouping similar instances into clusters according to some distance measures. The main idea is to put instances that are similar (that is, close to each other) into the same cluster, while keeping the dissimilar points (that is, the ones further apart from each other) in different clusters. An example of how clusters might look like is shown in the following diagram:

The clustering algorithms follow two fundamentally different approaches. The first is a hierarchical or agglomerative approach that first considers each point as its own cluster, and then iteratively merges the most similar clusters together. It stops when further merging reaches a predefined number of clusters, or if the clusters to be merged are spread over a large region.

The other approach is based on point assignment. First, initial cluster centers (that is, centroids) are estimated, for instance, randomly, and then, each point is assigned to the closest cluster, until all of the points are assigned. The most well known algorithm in this group is k-means clustering.

The k-means clustering either picks initial cluster centers as points that are as far as possible from one another, or (hierarchically) clusters a sample of data and picks a point that is the closest to the center of each of the k-clusters.

Supervised learning is the key concept behind such amazing things as voice recognition, email spam filtering, and face recognition in photos, and detecting credit card frauds. More formally, given a set, D, of learning examples described with features, X, the goal of supervised learning is to find a function that predicts a target variable, Y. The function, f ,that describes the relation between features X and class Y is called a model:

The general structure of supervised learning algorithms is defined by the following decisions (Hand et al., 2001):

1. Define the task
2. Decide on the machine learning algorithm, which introduces specific inductive bias; that is, and a priori assumptions that it makes regarding the target concept
3. Decide on the score or cost function, for instance, information gain, root mean square error, and so on
4. Decide on the optimization/search method to optimize the score function
5. Find a function that describes the relation between X and Y

Many decisions are already made for us by the type of the task and dataset that we have. In the following sections, we will take a closer look at the classification and regression methods and the corresponding score functions.

Classification can be applied when we deal with a discrete class, where the goal is to predict one of the mutually exclusive values in the target variable. An example would be credit scoring, where the final prediction is whether the person is credit liable or not. The most popular algorithms include decision trees, Naive Bayes classifiers, SVMs, neural networks, and ensemble methods.

Decision tree learning builds a classification tree, where each node corresponds to one of the attributes; edges correspond to a possible value (or intervals) of the attribute from which the node originates; and each leaf corresponds to a class label. A decision tree can be used to visually and explicitly represent the prediction model, which makes it a very transparent (white box) classifier. Notable algorithms are ID3 and C4.5, although many alternative implementations and improvements exist (for example, J48 in Weka).

Given a set of attribute values, a probabilistic classifier is able to predict a distribution over a set of classes, rather than an exact class. This can be used as a degree of certainty; that is, how sure the classifier is about its prediction. The most basic classifier is Naive Bayes, which happens to be the optimal classifier if, and only if, the attributes are conditionally independent. Unfortunately, this is extremely rare in practice.

There is an enormous subfield denoted as probabilistic graphical models, comprising hundreds of algorithms for example, Bayesian networks, dynamic Bayesian networks, hidden Markov models, and conditional random fields that can handle not only specific relationships between attributes, but also temporal dependencies. Kiran R Karkera wrote an excellent introductory book on this topic, Building Probabilistic Graphical Models with Python, Packt Publishing (2014), while Koller and Friedman published a comprehensive theory bible, Probabilistic Graphical Models, MIT Press (2009).

Any linear model can be turned into a non-linear model by applying the kernel trick to the model—replacing its features (predictors) by a kernel function. In other words, the kernel implicitly transforms our dataset into higher dimensions. The kernel trick leverages the fact that it is often easier to separate the instances in more dimensions. Algorithms capable of operating with kernels include the kernel perceptron, SVMs, Gaussian processes, PCA, canonical correlation analysis, ridge regression, spectral clustering, linear adaptive filters, and many others.

Artificial neural networks are inspired by the structure of biological neural networks and are capable of machine learning, as well as pattern recognition. They are commonly used for both regression and classification problems, comprising a wide variety of algorithms and variations for all manner of problem types. Some popular classification methods are perceptron, restricted Boltzmann machine (RBM), and deep belief networks.

Ensemble methods compose of a set of diverse weaker models to obtain better predictive performance. The individual models are trained separately and their predictions are then combined in some way to make the overall prediction. Ensembles, hence, contain multiple ways of modeling the data, which hopefully leads to better results. This is a very powerful class of techniques, and as such, it is very popular. This class includes boosting, bagging, AdaBoost, and random forest. The main differences among them are the type of weak learners that are to be combined and the ways in which to combine them.

Is our classifier doing well? Is this better than the other one? In classification, we count how many times we classify something right and wrong. Suppose there are two possible classification labels of yes and no, then there are four possible outcomes, as shown in the following table:

The four variables:

• True positive (hit): This indicates a yes instance correctly predicted as yes
• True negative (correct rejection): This indicates a no instance correctly predicted as no
• False positive (false alarm): This indicates a no instance predicted as yes
• False negative (miss): This indicates a yes instance predicted as no

The basic two performance measures of a classifier are, firstly, classification error:

And, secondly, classification accuracy is another performance measure, as shown here:

The main problem with these two measures is that they cannot handle unbalanced classes. Classifying whether a credit card transaction is an abuse or not is an example of a problem with unbalanced classes: there are 99.99% normal transactions and just a tiny percentage of abuses. The classifier that says that every transaction is a normal one is 99.99% accurate, but we are mainly interested in those few classifications that occur very rarely.

The solution is to use measures that don't involve true negatives. Two such measures are as follows:

• Precision: This is the proportion of positive examples correctly predicted as positive (TP) out of all examples predicted as positive (TP + FP):

• Recall: This is the proportion of positives examples correctly predicted as positive (TP) out of all positive examples (TP + FN):

It is common to combine the two and report the F-measure, which considers both precision and recall to calculate the score as a weighted average, where the score reaches its best value at 1 and worst at 0, as follows:

Most classification algorithms return a classification confidence denoted as f(X), which is, in turn, used to calculate the prediction. Following the credit card abuse example, a rule might look similar to the following:

The threshold determines the error rate and the true positive rate. The outcomes of all the possible threshold values can be plotted as receiver operating characteristics (ROC) as shown in the following diagram:

A random predictor is plotted with a red dashed line and a perfect predictor is plotted with a green dashed line. To compare whether the A classifier is better than C, we compare the area under the curve.

Most of the toolboxes provide all of the previous measures out of the box.

Regression deals with a continuous target variable, unlike classification, which works with a discrete target variable. For example, in order to forecast the outside temperature of the following few days, we would use regression, while classification will be used to predict whether it will rain or not. Generally speaking, regression is a process that estimates the relationship among features, that is, how varying a feature changes the target variable.

The most basic regression model assumes linear dependency between features and target variable. The model is often fitted using least squares approach, that is, the best model minimizes the squares of the errors. In many cases, linear regression is not able to model complex relations; for example, the following diagram shows four different sets of points having the same linear regression line. The upper-left model captures the general trend and can be considered as a proper model, whereas the bottom-left model fits points much better (except for one outlier, which should be carefully checked), and the upper and lower-right side linear models completely miss the underlying structure of the data and cannot be considered proper models:

Linear regression works when the dependent variable is continuous. If, however, the dependent variable is binary in nature, that is, 0 or 1, success or failure, yes or no, true or false, survived or died, and so on, then logistic regression is used instead. One such example is a clinical trial of drugs where the subject under study either responds to the drugs or does not respond. It is also used in fraud detection where the transaction is either a fraud or not fraud. Normally, a logistic function is used to measure the relationship between dependent and independent variables. It is seen as a Bernoulli distribution and, when plotted, looks similar to a curve in the shape of characters.

In regression, we predict numbers, Y, from input, X, and the predictions are usually wrong or not exact. The main question that we have to ask is: by how much? In other words, we want to measure the distance between the predicted and true values.

Mean squared error (MSE) is an average of the squared difference between the predicted and true values, as follows:

The measure is very sensitive to the outliers, for example, 99 exact predictions and 1 prediction off by 10 is scored the same as all predictions wrong by 1. Moreover, the measure is sensitive to the mean. Therefore, a relative squared error that compares the MSE of our predictor to the MSE of the mean predictor (which always predicts the mean value) is often used instead.

Mean absolute error (MAS) is an average of the absolute difference between the predicted and the true values, as follows:

The MAS is less sensitive to the outliers, but it is also sensitive to the mean and scale.

Correlation coefficient (CC) compares the average of prediction relative to the mean, multiplied by training values relative to the mean. If the number is negative, it means weak correlation; a positive number means strong correlation; and zero means no correlation. The correlation between true values X and predictions Y is defined as follows:

The CC measure is completely insensitive to the mean and scale and less sensitive to the outliers. It is able to capture the relative ordering, which makes it useful for ranking tasks, such as document relevance and gene expression.

Once the model is built, how do we know it will perform on new data? Is this model any good? To answer these questions, we'll first look into the model generalization, and then see how to get an estimate of the model performance on new data.

Predictor training can lead to models that are too complex or too simple. The model with low complexity (the leftmost models in the following diagram) can be as simple as predicting the most frequent or mean class value, while the model with high complexity (the rightmost models) can represent the training instances. Modes that are too rigid, shown on the left-hand side, cannot capture complex patterns; while models that are too flexible, shown on the right-hand side, fit to the noise in the training data. The main challenge is to select the appropriate learning algorithm and its parameters, so that the learned model will perform well on the new data (for example, the middle column):

The following diagram shows how errors in the training set decreases with model complexity. Simple rigid models underfit the data and have large errors. As model complexity increases, it describes the underlying structure of the training data better and, consequentially, the error decreases. If the model is too complex, it overfits the training data and its prediction error increases again:

Depending on the task complexity and data availability, we want to tune our classifiers toward more or less complex structures. Most learning algorithms allow such tuning, as follows:

• Regression: This is the order of the polynomial
• Naive Bayes: This is the number of the attributes
• Decision trees: This is the number of nodes in the tree—pruning confidence
• K-nearest neighbors: This is the number of neighbors—distance-based neighbor weights
• SVM: This is the kernel type; cost parameter
• Neural network: This is the number of neurons and hidden layers

With tuning, we want to minimize the generalization error; that is, how well the classifier performs on future data. Unfortunately, we can never compute the true generalization error; however, we can estimate it. Nevertheless, if the model performs well on the training data but performance is much worse on the test data, then the model most likely overfits.

To estimate the generalization error, we split our data into two parts: training data and testing data. A general rule of thumb is to split them by the training: testing ratio, that is, 70:30. We first train the predictor on the training data, then predict the values for the test data, and finally, compute the error, that is, the difference between the predicted and the true values. This gives us an estimate of the true generalization error.

The estimation is based on the two following assumptions: first, we assume that the test set is an unbiased sample from our dataset; and second, we assume that the actual new data will reassemble the distribution as our training and testing examples. The first assumption can be mitigated by cross-validation and stratification. Also, if it is scarce, one can't afford to leave out a considerable amount of data for a separate test set, as learning algorithms do not perform well if they don't receive enough data. In such cases, cross-validation is used instead.

Cross-validation splits the dataset into k sets of approximately the same size—for example, in the following diagram, into five sets. First, we use sets 2 to 5 for learning and set 1 for training. We then repeat the procedure five times, leaving out one set at a time for testing, and average the error over the five repetitions:

This way, we use all of the data for learning and testing as well, while avoiding using the same data to train and test a model.

An extreme example of cross-validation is the leave-one-out validation. In this case, the number of folds is equal to the number of instances; we learn on all but one instance, and then test the model on the omitted instance. We repeat this for all instances, so that each instance is used exactly once for the validation. This approach is recommended when we have a limited set of learning examples, for example, less than 50.

Stratification is a procedure to select a subset of instances in such a way that each fold roughly contains the same proportion of class values. When a class is continuous, the folds are selected so that the mean response value is approximately equal in all of the folds. Stratification can be applied along with cross-validation or separate training and test sets.

In this chapter, we refreshed our knowledge of machine learning basics. We revisited the workflow of applied machine learning and clarified the main tasks, methods, and algorithms. We learned about different types for regression and how to evaluate them. We also explored cross-validation and where it is applied.

In the next chapter, we will learn about Java libraries, the tasks that they can perform, and different platforms for machine learning.