Machine learning

Machine learning has been around for many years now and all social media users, at some point in time, have been consumers of Machine learning technology. One of the common examples is face recognition software, which is the capability to identify whether a digital photograph includes a given person. Today, Facebook users can see automatic suggestions to tag their friends in the digital photographs that are uploaded. Some cameras and software such as iPhoto also have this capability. There are many examples and use cases that will be discussed in more detail later in this chapter.

The following concept map represents the key aspects and semantics of Machine learning that will be covered throughout this chapter:

Definition

Let's start with defining what Machine learning is. There are many technical and functional definitions for Machine learning, and some of them are as follows:

	"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
	--Tom M. Mitchell

	"Machine learning is the training of a model from data that generalizes a decision against a performance measure."
	--Jason Brownlee

	"A branch of artificial intelligence in which a computer generates rules underlying or based on raw data that has been fed into it."
	--Dictionary.com

	"Machine learning is a scientific discipline that is concerned with the design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases."
	--Wikipedia

The preceding definitions are fascinating and relevant. They either have an algorithmic, statistical, or mathematical perspective.

Beyond these definitions, a single term or definition for Machine learning is the key to facilitating the definition of a problem-solving platform. Basically, it is a mechanism for pattern search and building intelligence into a machine to be able to learn, implying that it will be able to do better in the future from its own experience.

Drilling down a little more into what a pattern typically is, pattern search or pattern recognition is essentially the study of how machines perceive the environment, learn to discriminate behavior of interest from the rest, and be able to take reasonable decisions about categorizing the behavior. This is more often performed by humans. The goal is to foster accuracy, speed, and avoid the possibility of inappropriate use of the system.

Machine learning algorithms that are constructed this way handle building intelligence. Essentially, machines make sense of data in much the same way that humans do.

The primary goal of a Machine learning implementation is to develop a general purpose algorithm that solves a practical and focused problem. Some of the aspects that are important and need to be considered in this process include data, time, and space requirements. Most importantly, with the ability to be applied to a broad class of learning problems, the goal of a learning algorithm is to produce a result that is a rule and is as accurate as possible.

Another important aspect is the big data context; that is, Machine learning methods are known to be effective even in cases where insights need to be uncovered from datasets that are large, diverse, and rapidly changing. More on the large scale data aspect of Machine learning will be covered in Chapter 2, Machine Learning and Large-scale Datasets.

Core Concepts and Terminology

At the heart of Machine learning is knowing and using the data appropriately. This includes collecting the right data, cleansing the data, and processing the data using learning algorithms iteratively to build models using certain key features of data, and based on the hypotheses from these models, making predictions.

In this section, we will cover the standard nomenclature or terminology used in machine learning, starting from how to describe data, learning, modeling, algorithms, and specific machine learning tasks.

What is learning?

Now, let us look at the definition of "learning" in the context of Machine learning. In simple terms, historical data or observations are used to predict or derive actionable tasks. Very clearly, one mandate for an intelligent system is its ability to learn. The following are some considerations to define a learning problem:

• Provide a definition of what the learner should learn and the need for learning.
• Define the data requirements and the sources of the data.
• Define if the learner should operate on the dataset in entirety or a subset will do.

Before we plunge into understanding the internals of each learning type in the following sections, you need to understand the simple process that is followed to solve a learning problem, which involves building and validating models that solve a problem with maximum accuracy.

Tip

A model is nothing but an output from applying an algorithm to a dataset, and it is usually a representation of the data. We cover more on models in the later sections.

In general, for performing Machine learning, there are primarily two types of datasets required. The first dataset is usually manually prepared, where the input data and the expected output data are available and prepared. It is important that every piece of input data has an expected output data point available as this will be used in a supervised manner to build the rule. The second dataset is where we have the input data, and we are interested in predicting the expected output.

As a first step, the given data is segregated into three datasets: training, validation, and testing. There is no one hard rule on what percentage of data should be training, validation, and testing datasets. It can be 70-10-20, 60-30-10, 50-25-25, or any other values.

The training dataset refers to the data examples that are used to learn or build a classifier, for example. The validation dataset refers to the data examples that are verified against the built classifier and can help tune the accuracy of the output. The testing dataset refers to the data examples that help assess the performance of the classifier.

There are typically three phases for performing Machine learning:

• Phase 1—Training Phase: This is the phase where training data is used to train the model by pairing the given input with the expected output. The output of this phase is the learning model itself.
• Phase 2—Validation and Test Phase: This phase is to measure how good the learning model that has been trained is and estimate the model properties, such as error measures, recall, precision, and others. This phase uses a validation dataset, and the output is a sophisticated learning model.
• Phase 3—Application Phase: In this phase, the model is subject to the real-world data for which the results need to be derived.

The following figure depicts how learning can be applied to predict the behavior:

Data

Data forms the main source of learning in Machine learning. The data that is being referenced here can be in any format, can be received at any frequency, and can be of any size. When it comes to handling large datasets in the Machine learning context, there are some new techniques that have evolved and are being experimented with. There are also more big data aspects, including parallel processing, distributed storage, and execution. More on the large-scale aspects of data will be covered in the next chapter, including some unique differentiators.

When we think of data, dimensions come to mind. To start with, we have rows and columns when it comes to structured and unstructured data. This book will cover handling both structured and unstructured data in the machine learning context. In this section, we will cover the terminology related to data within the Machine learning context.

Term	Purpose or meaning in the context of Machine learning
Feature, attribute, field, or variable	This is a single column of data being referenced by the learning algorithms. Some features can be input to the learning algorithm, and some can be the outputs.
Instance	This is a single row of data in the dataset.
Feature vector or tuple	This is a list of features.
Dimension	This is a subset of attributes used to describe a property of data. For example, a date dimension consists of three attributes: day, month, and year.
Dataset	A collection of rows or instances is called a dataset. In the context of Machine learning, there are different types of datasets that are meant to be used for different purposes. An algorithm is run on different datasets at different stages to measure the accuracy of the model. There are three types of dataset: training, testing, and evaluation datasets. Any given comprehensive dataset is split into three categories of datasets and is usually in the following proportions: 60% training, 30% testing, and 10% evaluation.
a. Training Dataset	The training dataset is the dataset that is the base dataset against which the model is built or trained.
b. Testing Dataset	The testing dataset is the dataset that is used to validate the model built. This dataset is also referred to as a validating dataset.
c. Evaluation Dataset	The evaluation dataset is the dataset that is used for final verification of the model (and can be treated more as user acceptance testing).
Data Types	Attributes or features can have different data types. Some of the data types are listed here: Categorical (for example: young, old). Ordinal (for example: 0, 1). Numeric (for example: 1.3, 2.1, 3.2, and so on).
Coverage	The percentage of a dataset for which a prediction is made or the model is covered. This determines the confidence of the prediction model.

Labeled and unlabeled data

Data in the Machine learning context can either be labeled or unlabeled. Before we go deeper into the Machine learning basics, you need to understand this categorization, and what data is used when, as this terminology will be used throughout this book.

Unlabeled data is usually the raw form of the data. It consists of samples of natural or human-created artifacts. This category of data is easily available in abundance. For example, video streams, audio, photos, and tweets among others. This form of data usually has no explanation of the meaning attached.

The unlabeled data becomes labeled data the moment a meaning is attached. Here, we are talking about attaching a "tag" or "label" that is required, and is mandatory, to interpret and define the relevance. For example, labels for a photo can be the details of what it contains, such as animal, tree, college, and so on, or, in the context of an audio file, a political meeting, a farewell party, and so on. More often, the labels are mapped or defined by humans and are significantly more expensive to obtain than the unlabeled raw data.

The learning models can be applied to both labeled and unlabeled data. We can derive more accurate models using a combination of labeled and unlabeled datasets. The following diagram represents labeled and unlabeled data. Both triangles and bigger circles represent labeled data and small circles represent unlabeled data.

The application of labeled and unlabeled data is discussed in more detail in the following sections. You will see that supervised learning adopts labeled data and unsupervised learning adopts unlabeled data. Semi-supervised learning and deep learning techniques apply a combination of labeled and unlabeled data in a variety of ways to build accurate models.

Tasks

A task is a problem that the Machine learning algorithm is built to solve. It is important that we measure the performance on a task. The term "performance" in this context is nothing but the extent or confidence with which the problem is solved. Different algorithms when run on different datasets produce a different model. It is important that the models thus generated are not compared, and instead, the consistency of the results with different datasets and different models is measured.

Algorithms

After getting a clear understanding of the Machine learning problem at hand, the focus is on what data and algorithms are relevant or applicable. There are several algorithms available. These algorithms are either grouped by the learning subfields (such as supervised, unsupervised, reinforcement, semi-supervised, or deep) or the problem categories (such as Classification, Regression, Clustering or Optimization). These algorithms are applied iteratively on different datasets, and output models that evolve with new data are captured.

Models

Models are central to any Machine learning implementation. A model describes data that is observed in a system. Models are the output of algorithms applied to a dataset. In many cases, these models are applied to new datasets that help the models learn new behavior and also predict them. There is a vast range of machine learning algorithms that can be applied to a given problem. At a very high level, models are categorized as the following:

• Logical models
• Geometric models
• Probabilistic models

Logical models

Logical models are more algorithmic in nature and help us derive a set of rules by running the algorithms iteratively. A Decision tree is one such example:

Geometric models

Geometric models use geometric concepts such as lines, planes, and distances. These models usually operate, or can operate, on high volumes of data. Usually, linear transformations help compare different Machine learning methods:

Probabilistic models

Probabilistic models are statistical models that employ statistical techniques. These models are based on a strategy that defines the relationship between two variables. This relationship can be derived for sure as this involves using a random background process. In most cases, a subset of the overall data can be considered for processing:

Viagra	Lottery	P(Y= Spam (Viagra, lottery))	P(Y= ham (Viagra, lottery))
0	0	0.31	0.69
0	1	0.65	0.35
1	0	0.80	0.20
1	1	0.40	0.60

Data and inconsistencies in Machine learning

This section details all the possible data inconsistencies that may be encountered while implementing Machine learning projects, such as:

• Under-fitting
• Over-fitting
• Data instability
• Unpredictable future

Fortunately, there are some established processes in place today to address these inconsistencies. The following sections cover these inconsistencies.

Under-fitting

A model is said to be under-fitting when it doesn't take into consideration enough information to accurately model the actual data. For example, if only two points on an exponential curve are mapped, this possibly becomes a linear representation, but there could be a case where a pattern does not exist. In cases like these, we will see increasing errors and subsequently an inaccurate model. Also, in cases where the classifier is too rigid or is not complex enough, under-fitting is caused not just due to a lack of data, but can also be a result of incorrect modeling. For example, if the two classes form concentric circles and we try to fit a linear model, assuming they were linearly separable, this could potentially result in under-fitting.

The accuracy of the model is determined by a measure called "power" in the statistical world. If the dataset size is too small, we can never target an optimal solution.

Over-fitting

This case is just the opposite of the under-fitting case explained before. While too small a sample is not appropriate to define an optimal solution, a large dataset also runs the risk of having the model over-fit the data. Over-fitting usually occurs when the statistical model describes noise instead of describing the relationships. Elaborating on the preceding example in this context, let's say we have 500,000 data points. If the model ends up catering to accommodate all 500,000 data points, this becomes over-fitting. This will in effect mean that the model is memorizing the data. This model works well as long as the dataset does not have points outside the curve. A model that is over-fit demonstrates poor performance as minor fluctuations in data tend to be exaggerated. The primary reason for over-fitting also could be that the criterion used to train the model is different from the criterion used to judge the efficacy of the model. In simple terms, if the model memorizes the training data rather than learning, this situation is seen to occur more often.

Now, in the process of mitigating the problem of under-fitting the data, by giving it more data, this can in itself be a risk and end up in over-fitting. Considering that more data can mean more complexity and noise, we could potentially end up with a solution model that fits the current data at hand and nothing else, which makes it unusable. In the following graph, with the increasing model complexity and errors, the conditions for over-fit and under-fit are pointed out:

Data instability

Machine learning algorithms are usually robust to noise within the data. A problem will occur if the outliers are due to manual error or misinterpretation of the relevant data. This will result in a skewing of the data, which will ultimately end up in an incorrect model.

Therefore, there is a strong need to have a process to correct or handle human errors that can result in building an incorrect model.

Unpredictable data formats

Machine learning is meant to work with new data constantly coming into the system and learning from that data. Complexity will creep in when the new data entering the system comes in formats that are not supported by the machine learning system. It is now difficult to say if our models work well for the new data given the instability in the formats that we receive the data, unless there is a mechanism built to handle this.

Practical Machine learning examples

In this section, let's explore some real-world machine learning applications. We covered various examples within the introductory section of this chapter and we will now cover some domain-specific examples with a brief description of each problem.

For online and offline applications, some of the following examples can easily be guessed. In the chapters to follow, a subset of these examples will be picked to demonstrate the practical implementation aspects using suitable Machine learning algorithms.

Problem / problem Domain	Description
Spam detection	The problem statement here is to identify which e-mails are "spam". A Machine learning algorithm can categorize an e-mail to be marked as spam based on some rules that it builds using some key features of e-mail data. Once an e-mail is marked as spam, that e-mail is then moved to the spam folder and the rest are left in the inbox.
Credit card fraud detection	This is one of the recent problems that credit card firms need a solution for. Based on the usage patterns of the credit card by the consumer and the purchase behavior of the customer, the need is to identify any transaction that is not potentially made by the customer and mark them as fraudulent for necessary action to be taken.
Digit recognition	This is a very simple use case that requires the ability to group posts based on the zip code. This includes the need to interpret a handwritten numeric accurately and bucket the posts based on the zip code for faster processing.
Speech recognition	Automated call centers need this capability where a user's request on the phone is interpreted and mapped to one of the tasks for execution. The moment the user request can be mapped to a task, its execution can be automated. A model of this problem will allow a program to understand and make an attempt to fulfill that request. The iPhone with Siri has this capability.
Face detection	This is one of the key features that today's social media websites provide. This feature provides an ability to tag a person across many digital photographs. This gives aptitude to a group or categorizes the photographs by a person. Some cameras and software such as iPhoto have this capability.
Product recommendation or customer segmentation	This capability is found in almost all of the top online shopping websites today. Given a purchase history for a customer and a large inventory of products, the idea is to identify those products that the customer will most likely be interested in buying, thus motivating more product purchases. There are many online shopping and social websites that support this feature (for example: Amazon, Facebook, Google+, and many others). There are other cases like the ability to predict whether a trial version customer opts for the paid version of the product.
Stock trading	This means predicting stock performance based on the current past stock movement. This task is critical to financial analysts and helps provide decision support when buying and selling stocks.
Sentiment analysis	Many times, we find that the customers make decisions based on opinions shared by others. For example, we buy a product because it has received positive feedback from the majority of its users. Not only in commercial businesses as detailed earlier, but sentiment analysis is also being used by political strategists to gauge public opinion on policy announcements or campaign messages.

Types of learning problems

This section focuses on elaborating different learning problem categories. Machine learning algorithms are also classified under these learning problems. The following figure depicts various types of learning problems:

Classification

Classification is a way to identify a grouping technique for a given dataset in such a way that depending on a value of the target or output attribute, the entire dataset can be qualified to belong to a class. This technique helps in identifying the data behavior patterns. This is, in short, a discrimination mechanism.

For example, a sales manager needs help in identifying a prospective customer and wants to determine whether it is worth spending the effort and time the customer demands. The key input for the manager is the customer's data, and this case is commonly referred to as Total Lifetime Value (TLV).

We take the data and start plotting blindly on a graph (as shown in the following graph) with the x axis representing the total items purchased and the y axis representing the total money spent (in multiples of hundreds of dollars). Now we define the criteria to determine, for example, whether a customer is good or bad. In the following graph, all the customers who spend more than 800 dollars in a single purchase are categorized as good customers (note that this is a hypothetical example or analysis).

Now when new customer data comes in, the sales manager can plot the new customers on this graph and based on which side they fall, predict whether the customer is likely to be good or bad.

Tip

Note that classification need not always be binary (yes or no, male or female, good or bad, and so on) and any number of classifications can be defined (poor, below average, average, above average, good) based on the problem definition.

Clustering

In many cases, the data analyst is just given some data and is expected to unearth interesting patterns that may help derive intelligence. The main difference between this task and that of a classification is that in the classification problem, the business user specifies what he/she is looking for (a good customer or a bad customer, a success or a failure, and so on).

Let's now expand on the same example considered in the classification section. Here the patterns to classify the customers are identified without any target in mind or any prior classification, and unlike running a classification, the results may always not be the same (for example, depending on how the initial centroids are picked). An example modeling method for clustering is k-means clustering. More details on k-means clustering is covered in the next section and in detail in the following chapters.

In short, clustering is a classification analysis that does not start with a specific target in mind (good/bad, will buy/will not buy).

Forecasting, prediction or regression

Similar to classification, forecasting or prediction is also about identifying the way things would happen in the future. This information is derived from past experience or knowledge. In some cases, there is not enough data, and there is a need to define the future through regression. Forecasting and prediction results are always presented along with the degree of uncertainty or probability. This classification of the problem type is also called rule extraction.

Let's take an example here, an agricultural scientist working on a new crop that she developed. As a trial, this seed was planted at various altitudes and the yield was computed. The requirement here is to predict the yield of the crop given the altitude details (and some more related data points). The relationship between yield gained and the altitude is determined by plotting a graph between the parameters. An equation is noted that fits most of the data points, and in cases where data does not fit the curve, we can get rid of the data. This technique is called regression.

Simulation

In addition to all the techniques we defined until now, there might be situations where the data in context itself has many uncertainty. For example, an outsourcing manager is given a task and can estimate with experience that the task can be done by an identified team with certain skills in 2-4 hours.

Let's say the cost of input material may vary between $100-120 and the number of employees who come to work on any given day may be between 6 and 9. An analyst then estimates how much time the project might take. Solving such problems requires the simulation of a vast amount of alternatives.

Typically in forecasting, classification, and unsupervised learning, we are given data and we really do not know how the data is interconnected. There is no equation to describe one variable as a function of others.

Essentially, data scientists combine one or more of the preceding techniques to solve challenging problems, which are:

• Web search and information extraction
• Drug design
• Predicting capital market behavior
• Understanding customer behavior
• Designing robots

Optimization

Optimization, in simple terms, is a mechanism to make something better or define a context for a solution that makes it the best.

Considering a production scenario, let's assume there are two machines that produce the desired product but one machine requires more energy for high speed in production and lower raw materials while the other requires higher raw materials and less energy to produce the same output in the same time. It is important to understand the patterns in the output based on the variation in inputs; a combination that gives the highest profits would probably be the one the production manager would want to know. You, as an analyst, need to identify the best possible way to distribute the production between the machines that gives him the highest profit.

The following image shows the point of highest profit when a graph was plotted for various distribution options between the two machines. Identifying this point is the goal of this technique.

Unlike the case of simulations where there is uncertainty associated with the input data, in optimization we not only have access to data, but also have the information on the dependencies and relationships between data attributes.

One of the key concepts in Machine learning is a process called induction. The following learning subfields use the induction process to build models. Inductive learning is a reasoning process that uses the results of one experiment to run the next set of experiments and iteratively evolve a model from specific information.

The following figure depicts various subfields of Machine learning. These subfields are one of the ways the machine learning algorithms are classified.

Supervised learning

Supervised learning is all about operating to a known expectation and in this case, what needs to be analyzed from the data being defined. The input datasets in this context are also referred to as "labeled" datasets. Algorithms classified under this category focus on establishing a relationship between the input and output attributes, and use this relationship speculatively to generate an output for new input data points. In the preceding section, the example defined for the classification problem is also an example of supervised learning. Labeled data helps build reliable models but is usually expensive and limited.

When the input and output attributes of the data are known, the key in supervised learning is the mapping between the inputs to outputs. There are quite a few examples of these mappings, but the complicated function that links up the input and output attributes is not known. A supervised learning algorithm takes care of this linking, and given a large dataset of input/output pairs, these functions help predict the output for any new input value.

Unsupervised learning

In some of the learning problems, we do not have any specific target in mind to solve. In the earlier section, we discussed clustering, which is a classification analyses where we do not start with a specific target in mind (good/bad, will buy/will not buy) and is hence referred to as unsupervised analyses or learning. The goal in this case is to decipher the structure in the data against the build mapping between input and output attributes of data and, in fact, the output attributes are not defined. These learning algorithms operate on an "unlabeled" dataset for this reason.

Semi-supervised learning

Semi-supervised learning is about using both labeled and unlabeled data to learn models better. It is important that there are appropriate assumptions for the unlabeled data and any inappropriate assumptions can invalidate the model. Semi-supervised learning gets its motivation from the human way of learning.

Reinforcement learning

Reinforcement learning is learning that focuses on maximizing the rewards from the result. For example, while teaching toddlers new habits, rewarding them every time they follow instructions works very well. In fact, they figure out what behavior helps them earn rewards. This is reinforcement learning, and it is also called credit assessment learning.

The most important thing is that in reinforcement learning the model is additionally responsible for making decisions for which a periodic reward is received. The results in this case, unlike supervised learning, are not immediate and may require a sequence of steps to be executed before the final result is seen. Ideally, the algorithm will generate a sequence of decisions that helps achieve the highest reward or utility.

The goal in this learning technique is to measure the trade-offs effectively by exploring and exploiting the data. For example, when a person has to travel from a point A to point B, there will be many ways that include travelling by air, water, road or by walking, and there is significant value in considering this data by measuring the trade-offs for each of these options. Another important aspect is the significance of a delay in the rewards. How would this affect learning? For example, in games like chess, any delay in reward identification may change the result.

Deep learning

Deep learning is an area of Machine learning that focuses on unifying Machine learning with artificial intelligence. In terms of the relationship with artificial neural networks, this field is more of an advancement to artificial neural networks that work on large amounts of common data to derive practical insights. It deals with building more complex neural networks to solve problems classified under semi-supervised learning and operates on datasets that have little labeled data. Some Deep learning techniques are listed as follows:

• Convolutional Networks
• Restricted Boltzmann Machine (RBM)
• Deep Belief Networks (DBN)
• Stacked Autoencoders