Big Data, Data Mining, and Machine Learning

Value Creation for Business Leaders and Practitioners

Jared Dean

• Publisher: Wiley
• Year: 2014
• Print ISBN: 9781118618042
• eBook ISBN: 9781118920701
• Edition: 1
• Topic: Informatik
• Subtopic: Data Mining

With big data analytics comes big insights into profitability

Big data is big business. But having the data and the computational power to process it isn't nearly enough to produce meaningful results. Big Data, Data Mining, and Machine Learning: Value Creation for Business Leaders and Practitioners is a complete resource for technology and marketing executives looking to cut through the hype and produce real results that hit the bottom line. Providing an engaging, thorough overview of the current state of big data analytics and the growing trend toward high performance computing architectures, the book is a detail-driven look into how big data analytics can be leveraged to foster positive change and drive efficiency.

With continued exponential growth in data and ever more competitive markets, businesses must adapt quickly to gain every competitive advantage available. Big data analytics can serve as the linchpin for initiatives that drive business, but only if the underlying technology and analysis is fully understood and appreciated by engaged stakeholders. This book provides a view into the topic that executives, managers, and practitioners require, and includes:

• A complete overview of big data and its notable characteristics
• Details on high performance computing architectures for analytics, massively parallel processing (MPP), and in-memory databases
• Comprehensive coverage of data mining, text analytics, and machine learning algorithms
• A discussion of explanatory and predictive modeling, and how they can be applied to decision-making processes

Big Data, Data Mining, and Machine Learning provides technology and marketing executives with the complete resource that has been notably absent from the veritable libraries of published books on the topic. Take control of your organization's big data analytics to produce real results with a resource that is comprehensive in scope and light on hyperbole.

PART ONE
The Computing Environment

With data collection, “the sooner the better” is always the best answer.

— Marissa Mayer

Data mining is going through a significant shift with the volume, variety, value and velocity of data increasing significantly each year. The volume of data created is outpacing the amount of currently usable data to such a degree that most organizations do not know what value is in their data. At the same that time data mining is changing, hardware capabilities have also undergone dramatic changes. Just as data mining is not one thing but a collection of many steps, theories, and algorithms, hardware can be dissected into a number of components. The corresponding component changes are not always in sync with this increased demand in data mining, machine learning, and big analytical problems.

The four components of disk, memory, central processing unit, and network can be thought of as four legs of the hardware platform stool. To have a useful stool, all the legs must be of the same length or users will be frustrated, stand up, and walk away to find a better stool; so too must the hardware system for data mining be in balance in regard to the components to give users the best experience for their analytical problems.

Data mining on any scale cannot be done without specialized software. In order to explain the evolution and progression of the hardware, there needs to be a small amount of background on the traditional interaction between hardware and software. Data mining software packages are discussed in detail in Part One.

In the past, traditional data mining software was implemented by loading data into memory and running a single thread of execution over the data. The process was constrained by the amount of memory available and the speed of a processor. If the process could not fit entirely into memory, the process would fail. The single thread of execution also failed to take advantage of multicore servers unless multiple users were on the system at the same time.

The main reason we are seeing dramatic changes in data mining is related to the changes in storage technologies as well as computational capabilities. However, all software packages cannot take advantage of current hardware capacity. This is especially true of the distributed computing model. A careful evaluation should be made to ensure that algorithms are distributed and effectively leveraging all the computing power available to you.

CHAPTER 1
Hardware

I am often asked what the best hardware configuration is for doing data mining. The only appropriate answer for this type of question is that it depends on what you are trying to do. There are a number of considerations to be weighed when deciding how to build an appropriate computing environment for your big data analytics.

STORAGE (DISK)

Storage of data is usually the first thing that comes to mind when the topic of big data is mentioned. It is the storage of data that allows us to keep a record of history so that it can be used to tell us what will likely happen in the future.

A traditional hard drive is made up of platters which are actual disks coated in a magnetized film that allow the encoding of 1s and 0s that make up data. The spindles that turn the vertically stacked platters are a critical part of rating hard drives because the spindles determine how fast the platters can spin and thus how fast the data can be read and written. Each platter has a single drive head; they both move in unison so that only one drive head is reading from a particular platter.

This mechanical operation is very precise and also very slow compared to the other components of the computer. It can be a large contributor to the time required to solve high-performance data mining problems.

To combat the weakness of disk speeds, disk arrays¹ became widely available, and they provide higher throughput. The maximum throughput of a disk array to a single system from external storage subsystems is in the range of 1 to 6 gigabytes (GB) per second (a speedup of 10 to 50 times in data access rates).

Another change in disk drives as a response to the big data era is that their capacity has increased 50% to 100% per year in the last 10 years. In addition, prices for disk arrays have remained nearly constant, which means the price per terabyte (TB) has decreased by half per year.

This increase in disk drive capacity has not been matched by the ability to transfer data to/from the disk drive, which has increased by only 15% to 20% per year. To illustrate this, in 2008, the typical server drive was 500 GB and had a data transfer rate of 98 megabytes per second (MB/sec). The entire disk could be transferred in about 85 minutes (500 GB = 500,000 MB/98 MB/sec). In 2013, there were 4 TB disks that have a transfer rate of 150 MB/sec, but it would take about 440 minutes to transfer the entire disk. When this is considered in light of the amount of data doubling every few years, the problem is obvious. Faster disks are needed.

Solid state devices (SSDs) are disk drives without a disk or any moving parts. They can be thought of as stable memory, and their data read rates can easily exceed 450 MB/sec. For moderate-size data mining environments, SSDs and their superior throughput rates can dramatically change the time to solution. SSD arrays are also available, but SSDs still cost significantly more per unit of capacity than hard disk drives (HDDs). SSD arrays are limited by the same external storage bandwidth as HDD arrays. So although SSDs can solve the data mining problem by reducing the overall time to read and write the data, converting all storage to SSD might be cost prohibitive. In this case, hybrid strategies that use different types of devices are needed.

Another consideration is the size of disk drives that are purchased for analytical workloads. Smaller disks have faster access times, and there can be advantages in the parallel disk access that comes from multiple disks reading data at the same time for the same problem. This is an advantage only if the software can take advantage of this type of disk drive configuration.

Historically, only some analytical software was capable of using additional storage to augment memory by writing intermediate results to disk storage. This extended the size of problem that could be solved but caused run times to go up. Run times rose not just because of the additional data load but also due to the slower access of reading intermediate results from disk instead of reading them from memory. For a typical desktop or small server system, data access to storage devices, particularly writing to storage devices, is painfully slow. A single thread of execution for an analytic process can easily consume 100 MB/sec, and the dominant type of data access is sequential read or write. A typical high-end workstation has a15K RPM SAS drive; the drive spins at 15,000 revolutions per minute and uses the SAS technology to read and write data at a rate of 100 to 150 MB/sec. This means that one or two cores can consume all of the disk bandwidth available. It also means that on a modern system with many cores, a large percentage of the central processing unit (CPU) resources will be idle for many data mining activities; this is not a lack of needed computation resources but the mismatch that exists among disk, memory, and CPU.

CENTRAL PROCESSING UNIT

The term “CPU” has had two meanings in computer hardware. CPU is used to refer to the plastic and steel case that holds all the essential elements of a computer. This includes the power supply, motherboard, peripheral cards, and so on. The other meaning of CPU is the processing chip located inside the plastic and steel box. In this book, CPU refers to the chip.

The speed of the CPU saw dramatic improvements in the 1980s and 1990s. CPU speed was increasing at such a rate that single threaded software applications would run almost twice as fast on new CPU versions as they became available. The CPU speedup was described by Gordon Moore, cofounder of Intel, in the famous Moore’s law, which is an observation that the number of transistors and integrated circuits that are able to be put in a given area doubles every two years and therefore instructions can be executed at twice the speed. This trend in doubling CPU speed continued into the 1990s, when Intel engineers observed that if the doubling trend continued, the heat that would be emitted from these chips would be as hot as the sun by 2010. In the early 2000s, the Moore’s law free lunch was over, at least in terms of processing speed. Processor speeds (frequencies) stalled, and computer companies sought new ways to increase performance. Vector units, present in limited form in x86 since the Pentium MMX instructions, were increasingly important to attaining performance and gained additional features, such as single- and then double-precision floating point.

In the early 2000s, then, chip manufacturers also turned to adding extra threads of execution into their chips. These multicore chips were scaled-down versions of the multiprocessor supercomputers, with the cores sharing resources such as cache memory. The number of cores located on a single chip has increased over time; today many server machines offer two six-core CPUs.

In comparison to hard disk data access, CPU access to memory is faster than a speeding bullet; the typical access is in the range of 10 to 30 GB/sec. All other components of the computer are racing to keep up with the CPU.

Graphical Processing Unit

The graphical processing unit (GPU) has gotten considerable publicity as an unused computing resource that could reduce the run times of data mining and other analytical problems by parallelizing the computations. The GPU is already found in every desktop computer in the world.

In the early 2000s, GPUs got into the computing game. Graphics processing has evolved considerably from early text-only displays of the first desktop computers. This quest for better graphics has been driven by industry needs for visualization tools. One example is engineers using three-dimensional (3D) computer-aided design (CAD) software to create prototypes of new designs prior to ever building them. An even bigger driver of GPU computing has been the consumer video game industry, which has seen price and performance trends similar to the rest of the consumer computing industry. The relentless drive to higher performance at lower cost has given the average user unheard-of performance both on the CPU and the GPU.

Three-dimensional graphics processing must process millions or billions of 3D triangles in 3D scenes multiple times per second to create animation. Placing and coloring all of these triangles in their 3D environment req...

Table of contents

• Cover
• Epigraph
• Series
• Titlepage
• Copyright
• Dedication
• Foreword
• Preface
• Acknowledgments
• Introduction
• Part One The Computing Environment
• Part Two Turning Data into Business Value
• Part Three Success Stories of Putting It All Together
• About the Author
• Appendix Nike+ Fuelband Script to Retrieve Information
• References
• Index
• End User License Agreement