*450*
*75*
*2MB*

*English*
*Pages [263]*
*Year 2015*

- Author / Uploaded
- Scott Stevens

Topic Science & Mathematics

Subtopic Mathematics

Mathematical Decision Making: Predictive Models and Optimization Course Guidebook Professor Scott P. Stevens James Madison University

PUBLISHED BY: THE GREAT COURSES Corporate Headquarters 4840 Westfields Boulevard, Suite 500 Chantilly, Virginia 20151-2299 Phone: 1-800-832-2412 Fax: 703-378-3819 www.thegreatcourses.com

Copyright © The Teaching Company, 2015

Printed in the United States of America This book is in copyright. All rights reserved. Without limiting the rights under copyright reserved above, no part of this publication may be reproduced, stored in or introduced into a retrieval system, or transmitted, in any form, or by any means (electronic, mechanical, photocopying, recording, or otherwise), without the prior written permission of The Teaching Company.

Scott P. Stevens, Ph.D. Professor of Computer Information Systems and Business Analytics James Madison University

P

rofessor Scott P. Stevens is a Professor of Computer Information Systems and Business Analytics at James Madison University (JMU) in Harrisonburg, Virginia. In 1979, he received B.S. degrees in both Mathematics and Physics from The Pennsylvania State University, where KH ZDV ¿UVW LQ KLV JUDGXDWLQJ FODVV LQ WKH &ROOHJH RI 6FLHQFH %HWZHHQ completing his undergraduate work and entering a doctoral program, Professor Stevens worked for Burroughs Corporation (now Unisys) in the Advanced Development Organization. Among other projects, he contributed to a proposal to NASA for the Numerical Aerodynamic Simulation Facility, a computerized wind tunnel that could be used to test aeronautical designs without building physical models and to create atmospheric weather models better than those available at the time. In 1987, Professor Stevens received his Ph.D. in Mathematics from The Pennsylvania State University, working under the direction of Torrence Parsons and, later, George E. Andrews, the world’s leading expert in the study of integer partitions. Professor Stevens’s research interests include analytics, combinatorics, graph theory, game theory, statistics, and the teaching of quantitative material. In collaboration with his JMU colleagues, he has published articles on a wide range of topics, including neural network prediction of survival in blunt-injured trauma patients; the effect of private school competition on public schools; standards of ethical computer usage in different countries; automatic data collection in business; the teaching of statistics and linear programming; and optimization of the purchase, transportation, and deliverability of natural gas from the Gulf of Mexico. His publications have appeared in a number of conference proceedings, as well as in the European i

Journal of Operational Research; the International Journal of Operations & Production Management; Political Research Quarterly; Omega: The International Journal of Management Science; Neural Computing & Applications; INFORMS Transactions on Education; and the Decision Sciences Journal of Innovative Education. 3URIHVVRU6WHYHQVKDVDFWHGDVDFRQVXOWDQWIRUDQXPEHURI¿UPVLQFOXGLQJ Corning Incorporated, C&P Telephone, and Globaltec. He is a member of the Institute for Operations Research and the Management Sciences and the Alpha Kappa Psi business fraternity. Professor Stevens’s primary professional focus since joining JMU in 1985 has been his deep commitment to excellence in teaching. He was the 1999 recipient of the Carl Harter Distinguished Teacher Award, JMU’s highest teaching award. He also has been recognized as an outstanding teacher ¿YH WLPHV LQ WKH XQLYHUVLW\¶V XQGHUJUDGXDWH EXVLQHVV SURJUDP DQG RQFH LQ its M.B.A. program. His teaching interests are wide and include analytics, statistics, game theory, physics, calculus, and the history of science. Much of his recent research focuses on the more effective delivery of mathematical concepts to students. Professor Stevens’s previous Great Course is Games People Play: Game Theory in Life, Business, and BeyondŶ

ii

Table of Contents

INTRODUCTION Professor Biography ............................................................................i Course Scope .....................................................................................1 LECTURE GUIDES LECTURE 1 The Operations Research Superhighway...........................................4 LECTURE 2 Forecasting with Simple Linear Regression .....................................13 LECTURE 3 Nonlinear Trends and Multiple Regression.......................................24 LECTURE 4 Time Series Forecasting ...................................................................31 LECTURE 5 Data Mining—Exploration and Prediction .........................................40 LECTURE 6 'DWD0LQLQJIRU$I¿QLW\DQG&OXVWHULQJ .............................................49 LECTURE 7 Optimization—Goals, Decisions, and Constraints ............................57 LECTURE 8 Linear Programming and Optimal Network Flow ..............................64 LECTURE 9 Scheduling and Multiperiod Planning ...............................................72 LECTURE 10 Visualizing Solutions to Linear Programs .........................................79 iii

Table of Contents LECTURE 11 Solving Linear Programs in a Spreadsheet ......................................88 LECTURE 12 Sensitivity Analysis—Trust the Answer? ...........................................98 LECTURE 13 Integer Programming—All or Nothing.............................................106 LECTURE 14 :KHUH,VWKH(I¿FLHQF\)URQWLHU" ................................................... 116 LECTURE 15 Programs with Multiple Goals .........................................................128 LECTURE 16 Optimization in a Nonlinear Landscape ..........................................137 LECTURE 17 Nonlinear Models—Best Location, Best Pricing .............................144 LECTURE 18 Randomness, Probability, and Expectation ....................................152 LECTURE 19 Decision Trees—Which Scenario Is Best? .....................................161 LECTURE 20 Bayesian Analysis of New Information ...........................................171 LECTURE 21 Markov Models—How a Random Walk Evolves ............................178 LECTURE 22 Queuing—Why Waiting Lines Work or Fail ....................................188 LECTURE 23 Monte Carlo Simulation for a Better Job Bid ..................................197

iv

Table of Contents LECTURE 24 Stochastic Optimization and Risk ...................................................210 SUPPLEMENTAL MATERIAL Entering Linear Programs into a Spreadsheet ...............................221 Glossary .........................................................................................230 Bibliography ....................................................................................250

v

vi

Mathematical Decision Making: Predictive Models and Optimization Scope:

P

eople have an excellent track record for solving problems that are small and familiar, but today’s world includes an ever-increasing number of situations that are complicated and unfamiliar. How can decision makers—individuals, organizations in the public or private sectors, or nations—grapple with these often-crucial concerns? In many cases, the tools they’re choosing are mathematical ones. Mathematical decision making is a collection of quantitative techniques that is intended to cut through irrelevant information to the heart of a problem, and then it uses powerful tools to investigate that problem in detail, leading to a good or even optimal solution. Such a problem-solving approach used to be the province only of the mathematician, the statistician, or the operations research professional. All RI WKLV FKDQJHG ZLWK WZR WHFKQRORJLFDO EUHDNWKURXJKV ERWK LQ WKH ¿HOG RI computing: automatic data collection and cheap, readily available computing power. Automatic data collection (and the subsequent storage of that data) often provides the analyst with the raw information that he or she needs. The universality of cheap computing power means that analytical techniques can be practically applied to much larger problems than was the case in the past. Even more importantly, many powerful mathematical techniques can now be executed much more easily in a computer environment—even a personal computer environment—and are usable by those who lack a professional’s knowledge of their intricacies. The intelligent amateur, with a bit of guidance, can now use mathematical techniques to address many more of the complicated or unfamiliar problems faced by organizations large and small. It is with this goal that this course was created. The purpose of this course is to introduce you to the most important prediction and optimization techniques—which include some aspects of statistics and data mining—especially those arising in operations research (or operational research). We begin each topic by developing a clear intuition of the purpose 1

of the technique and the way it works. Then, we apply it to a problem in a step-by-step approach. When this involves using a computer, as often it does, we keep it accessible. Our work can be done in a spreadsheet environment, VXFKDV2SHQ2I¿FH¶V&DOFZKLFKLVIUHHO\GLVWULEXWDEOH RU0LFURVRIW2I¿FH¶V Excel. This has two advantages. First, it allows you to see our progress each step of the way. Second, it gives you easy access to an environment where you can try out what we’re examining on your own. Along the way, we explore many real-world situations where various prediction and optimization techniques have been applied—by individuals, by companies, by agencies in the public sector, and by nations all over the world. Just as there are many kinds of problems to be solved, there are many techniques for addressing them. These tools can broadly be divided into predictive models and mathematical optimization. Predictive models allow us to take what we already know about the behavior of a system and use it to predict how that system will behave in new circumstances. Regression, for example, allows us to explore the nature of the interdependence of related quantities, identifying those ones that are most XVHIXOLQSUHGLFWLQJWKHRQHWKDWSDUWLFXODUO\KROGVRXULQWHUHVWVXFKDVSUR¿W Sometimes, what we know about a system comes from its historical behavior, and we want to extrapolate from that. Time series forecasting allows us to take historical data as a guide, using it to predict what will happen next and informing us how much we can trust that prediction.

Scope

7KHÀRRGRIGDWDUHDGLO\DYDLODEOHWRWKHPRGHUQLQYHVWLJDWRUJHQHUDWHVDQHZ kind of challenge: how to sift through those gigabytes of raw information and identify the meaningful patterns hidden within them. This is the province of data mining, a hot topic with broad applications—from online searches to advertising strategies and from recognizing spam to identifying deadly genes in DNA. But making informed predictions is only half of mathematical decision making. We also look closely at optimization problems, where the goal is WR ¿QG D EHVW DQVZHU WR D JLYHQ SUREOHP 6XFFHVV LQ WKLV UHJDUG GHSHQGV crucially on creating a model of the situation in a mathematical form, and 2

we’ll spend considerable time on this important step. As we’ll discover, some optimization problems are amazingly easy to solve while others are much more challenging, even for a computer. We’ll determine what makes the difference and how we can address the obstacles. Because our input data isn’t always perfect, we’ll also analyze how sensitive our answers are to changes in those inputs. But uncertainty can extend beyond unreliable inputs. Much of life involves unpredictable events, so we develop a variety of techniques intended to help us make good decisions in the face of that uncertainty. Decision trees allow us to analyze events that unfold sequentially through time and evaluate future scenarios, which often involve uncertainty. Bayesian analysis allows us to update our probabilities of upcoming events in light of more recent information. Markov analysis allows us to model the evolution of a chance process over time. Queuing theory analyzes the behavior of waiting lines— not only for customers, but also for products, services, and Internet data packets. Monte Carlo simulation allows us to create a realistic model of an environment and then use a computer to create thousands of possible futures for it, giving us insights on how we can expect things to unfold. Finally, stochastic optimization brings optimization techniques to bear even in the face of uncertainty, in effect uniting the entire toolkit of deterministic and probabilistic approaches to mathematical decision making presented in this course. Mathematical decision making goes under many different names, depending on the application: operations research, mathematical optimization, analytics, business intelligence, management science, and others. But no matter what you call it, the result is a set of tools to understand any organization’s SUREOHPVPRUHFOHDUO\WRDSSURDFKWKHLUVROXWLRQVPRUHVHQVLEO\DQGWR¿QG good answers to them more consistently. This course will teach you how some fairly simple math and a little bit of typing in a spreadsheet can be SDUOD\HGLQWRDVXUSULVLQJDPRXQWRISUREOHPVROYLQJSRZHUŶ

3

The Operations Research Superhighway Lecture 1

T

Lecture 1: The Operations Research Superhighway

KLV FRXUVH LV DOO DERXW WKH FRQÀXHQFH RI PDWKHPDWLFDO WRROV and computational power. Taken as a whole, the discipline of mathematical decision making has a variety of names, including operational research, operations research, management science, quantitative management, and analytics. But its purpose is singular: to apply quantitative methods to help people, businesses, governments, public services, military RUJDQL]DWLRQV HYHQW RUJDQL]HUV DQG ¿QDQFLDO LQYHVWRUV ¿QG ZD\V WR GR what they do better. In this lecture, you will be introduced to the topic of operations research. What Is Operations Research? z Operations research is an umbrella term that encompasses many powerful techniques. Operations research applies a variety of mathematical techniques to real-world problems. It leverages those techniques by taking advantage of today’s computational power. And, if successful, it comes up with an implementation strategy to make the situation better. This course is about some of the most important and most widely applicable ways that that gets done: through predictive models and mathematical optimization.

4

z

In broad terms, predictive models allow us to take what we already know about the behavior of a system and use it to predict how that system will behave in new circumstances. Often, what we know about a system comes from its historical behavior, and we want to extrapolate from that.

z

Sometimes, it’s not history that allows us to make predictions but, instead, what we know about how the pieces of the system ¿WWRJHWKHU&RPSOH[EHKDYLRUFDQHPHUJHIURPWKHLQWHUDFWLRQRI even simple parts. From there, we can investigate the possibilities— and probabilities.

z

But making informed predictions is only half of what this course is about. We’ll also be looking closely at optimization and the WRROVWRDFFRPSOLVKLW2SWLPL]DWLRQPHDQV¿QGLQJWKHEHVWDQVZHU possible to a problem. And the situation can change before the best answer that you found has to be scrapped. There are a variety of optimization techniques, and some optimization questions are much harder to solve than others.

z

Mathematical decision making offers a different way of thinking about problems. This way of looking at problems goes all the ZD\ EDFN WR WKH ULVH RI WKH VFLHQWL¿F DSSURDFK²LQ SDUWLFXODU investigating the world not only qualitatively but quantitatively. That change turned alchemy into chemistry, natural philosophy into physics and biology, astrology into astronomy, and folk remedies into medicine.

z

It took a lot longer for this mindset to make its way from science DQGHQJLQHHULQJLQWRRWKHU¿HOGVVXFKDVEXVLQHVVDQGSXEOLFSROLF\ In the 1830s, Charles Babbage, the pioneer in early computing machines, expounded what today is called the Babbage principle— namely, the idea that highly skilled, high-cost laborers should not be “wasting” their time on work that lower-skilled, lower-cost laborers could be doing.

z

,QWKHVWKLVLGHDEHFDPHSDUWRI)UHGULFN7D\ORU¶VVFLHQWL¿F management, which attempted to apply the principles of science WRPDQXIDFWXULQJZRUNÀRZ+LVDSSURDFKIRFXVHGRQVXFKPDWWHUV DV HI¿FLHQF\ NQRZOHGJH WUDQVIHU DQDO\VLV DQG PDVV SURGXFWLRQ Tools of statistical analysis began to be applied to business.

z

Then, Henry Ford took the idea of mass production, coupled it with interchangeable parts, and developed the assembly line system at his Ford Motor Company. The result was a company that, in the early 20th century, paid high wages to its workers and still sold an affordable automobile.

5

Lecture 1: The Operations Research Superhighway

z

But most historians set the real start of operations research in Britain in 1937 during the perilous days leading up to World War II— VSHFL¿FDOO\WKH%DZGVH\5HVHDUFK6WDWLRQQHDU6XIIRON,WZDVWKH center of radar research and development in Britain at the time. It ZDVDOVRWKHORFDWLRQRIWKH¿UVWUDGDUWRZHULQZKDWEHFDPH%ULWDLQ¶V essential early-warning system against the German Luftwaffe.

z

A. P. Rowe was the station superintendent in 1937, and he wanted to investigate how the system might be improved. Rowe not only assessed the equipment, but he also studied the behavior of the operators of the equipment, who were, after all, soldiers acting as technicians. The results allowed Britain to improve the performance RI ERWK PHQ DQG PDFKLQHV 5RZH¶V ZRUN DOVR LGHQWL¿HG VRPH previously unnoticed weaknesses in the system.

z

This analytical approach was dubbed “operational research” by the British, and it quickly spread to other branches of their military and to the armed forces of other allied countries.

Computing Power z Operational research—or, as it came to be known in the United States, operations research—was useful throughout the war. It doubled the on-target bomb rate for B-29s attacking Japan. It increased U-boat hunting kill rates by about a factor of 10. Most RI WKLV DQG RWKHU ZRUN ZDV FODVVL¿HG GXULQJ WKH ZDU \HDUV 6R it wasn’t until after the war that people started turning a serious eye toward what operational research could do in other areas. And the real move in that direction started in the 1950s, with the introduction of the electronic computer. z

6

Until the advent of the modern computer, even if we knew how to solve a problem from a practical standpoint, it was often just too much work. Weather forecasting, for example, had some mathematical techniques available from the 1920s, but it was impossible to reasonably compute the predictions of the models before the actual weather occurred.

z

Computers changed that in a big way. And the opportunities have only accelerated in more recent decades. Gordon E. Moore, FRIRXQGHU RI ,QWHO ¿UVW VXJJHVWHG LQ ZKDW KDV VLQFH FRPH to be known as Moore’s law: that transistor chip count on an integrated circuit doubles about every two years. Many things that we care about, such as processor speed and memory capacity, grow along with it. Over more than 50 years, the law has continued to be remarkably accurate.

z

It’s hard to get a grip on how much growth that kind of doubling implies. Moore’s law accurately predicted that the number of chips on an integrated circuit in 2011 was about 8 million times as high as it was in 1965. That’s roughly the difference between taking a single step and walking from Albany, Maine, to Seattle, Washington, by way of Houston and Los Angeles. All of that power was now available to individuals and companies at an affordable price.

Mathematical Decision-Making Techniques z Once we have the complicated and important problems, like it or not, along with the computing power, the last piece of the puzzle is the mathematical decision-making techniques that allow us to better understand the problem and put all that computational power to work. z

7R GR WKLV ¿UVW \RX KDYH WR GHFLGH ZKDW \RX¶UH WU\LQJ WR accomplish. Then, you have to get the data that’s relevant to the problem at hand. Data collection and cleansing can always be a challenge, but the computer age makes it easier than ever before. So much information is automatically collected, and much of it can be retrieved with a few keystrokes.

z

But then comes what is perhaps the key step. The problem lives in the real world, but in order to use the powerful synergy of mathematics and computers, it has to be transported into a new, more abstract world. The problem is translated from the English that we use to describe it to each other into the language of

7

Lecture 1: The Operations Research Superhighway

mathematics. Mathematical language isn’t suited to describe everything, but what it can capture it does with unparalleled precision and stunning economy. z

Once you’ve succeeded in creating your translation—once you have modeled the problem—you look for patterns. You try to see how this new problem is like ones you’ve seen before and then apply your experience with them to it.

z

But when an operations researcher thinks about what other problems are similar to the current one, he or she is thinking about, most of all, the mathematical formulation, not the real-world context. In daily life, you might have useful categories like business, medicine, or engineering, but relying on these categories in operations research is as sensible as thinking that if you know how to buy a car, then you know how to make one, because both tasks deal with cars.

z

In operations research, the categorization of a problem depends on the mathematical character of the problem. The industry from which it comes only matters in helping to specify the mathematical character of the problem correctly.

Modeling and Formulation z The translation of a problem from English to math involves modeling and formulation. An important way that we can classify problems is as either stochastic or deterministic. Stochastic problems involve random elements; deterministic problems don’t.

8

z

Many problems ultimately have both deterministic and stochastic elements, so it’s helpful to begin this course with some statistics and data mining to get a sense of that combination. Both topics DUH ¿HOGV LQ WKHLU RZQ ULJKW WKDW RIWHQ SOD\ LPSRUWDQW UROHV LQ operations research.

z

Many deterministic operations research problems focus on optimization. For problems that are simple or on a small scale, the optimal solution may be obvious. But as the scale or complexity

of the problem increases, the number of possible courses of action tends to explode. And experience shows that seat-of-the-pants decision making can often result in terrible strategies. z

But once the problem is translated into mathematics, we can apply WKHIXOOSRZHURIWKDWGLVFLSOLQHWR¿QGLQJLWVEHVWDQVZHU,QDUHDO VHQVHWKHVHSUREOHPVFDQRIWHQEHWKRXJKWRIDV¿QGLQJWKHKLJKHVW or lowest point in some mathematical landscape. And how we do this is going to depend on the topography of that landscape. It’s easier to navigate a pasture than a glacial moraine. It’s also easier to ¿QG\RXUZD\WKURXJKRSHQFRXQWU\VLGHWKDQWKURXJKDODQGVFDSH crisscrossed by fences.

z

&DOFXOXV KHOSV ZLWK ¿QGLQJ KLJKHVW DQG ORZHVW SRLQWV DW OHDVW when the landscape is rolling hills and the fences are well behaved, or non-existent. But in calculus, we tend to have complicated functions and simple boundary conditions. For many of the practical problems we’ll explore in this course through linear programming, we have exactly the opposite: simple functions but complicated boundary conditions.

z

In fact, calculus tends to be useless and irrelevant for linear functions, both because the derivatives involved are all constants and because the optimum of a linear function is always on the boundary of its domain, never where the derivative is zero. So, we’re going to focus on other ways of approaching optimization problems—ways that don’t require a considerable background in calculus and that are better at handling problems with cliffs and fences.

z

These deterministic techniques often allow companies to use computer power to solve in minutes problems that would take hours or days to sort out on our own. But what about more sizeable uncertainty? As soon as the situation that you’re facing involves a random process, you’re probably not going to be able to guarantee WKDW\RX¶OO¿QGWKHEHVWDQVZHUWRWKHVLWXDWLRQ²DWOHDVWQRWD³EHVW answer” in the sense that we mean it for deterministic problems.

9

z

For example, given the opportunity to buy a lottery ticket, the best strategy is to buy it if it’s a winning ticket and don’t buy it if it’s not. But, of course, you don’t know whether it’s a winner or a loser at the time you’re deciding on the purchase. So, we have to come up with a different way to measure the quality of our decisions when we’re dealing with random processes. And we’ll need different techniques, including probability, Bayesian statistics, Markov analysis, and simulation.

Important Terms derivative: The derivative of a function is itself a function, one that HVVHQWLDOO\VSHFL¿HVWKHVORSHRIWKHRULJLQDOIXQFWLRQDWHDFKSRLQWDWZKLFK LW LV GH¿QHG )RU IXQFWLRQV RI PRUH WKDQ RQH YDULDEOH WKH FRQFHSW RI D derivative is captured by the vector quantity of the gradient.

Lecture 1: The Operations Research Superhighway

deterministic: Involving no random elements. For a deterministic problem, the same inputs always generate the same outputs. Contrast to stochastic. model $ VLPSOL¿HG UHSUHVHQWDWLRQ RI D VLWXDWLRQ WKDW FDSWXUHV WKH NH\ elements of the situation and the relationships among those elements. Moore’s law: Formulated by Intel founder Gordon Moore in 1965, it is the prediction that the number of transistors on an integrated circuit doubles roughly every two years. To date, it’s been remarkably accurate. operations research: The general term for the application of quantitative WHFKQLTXHVWR¿QGJRRGRURSWLPDOVROXWLRQVWRUHDOZRUOGSUREOHPV2IWHQ called operational research in the United Kingdom. When applied to business problems, it may be referred to as management science, business analytics, or quantitative management. optimization: Finding the best answer to a given problem. The best answer is termed “optimal.”

10

optimum: The best answer. The best answer among all possible solutions is a global optimum. An answer that is the best of all points in its immediate vicinity is a local optimum. Thus, in considering the heights of points in a mountain range, each mountain peak is a local maximum, but the top of the tallest mountain is the global maximum. stochastic: Involving random elements. Identical inputs may generate differing outputs. Contrast to deterministic.

Suggested Reading Budiansky, Blackett’s War. Gass and Arjang, An Annotated Timeline of Operations Research. Horner and List, “Armed with O.R.” Yu, Argüello, Song, McCowan, and White, “A New Era for Crew Recovery at Continental Airlines.”

Questions and Comments 1. Suppose that you decide to do your holiday shopping online. You have a complete list of the presents desired by your friends and family as well as access to the inventory, prices, and shipping costs for each online site. How could you characterize your task as a deterministic optimization problem? What real-world complications may turn your problem from a deterministic problem into a stochastic one? Answer: The most obvious goal is to minimize total money spent, but it is by no means the only possibility. If you are feeling generous, you might wish to maximize number of presents bought, maximize number of people for whom you give presents, and so on. You’ll face some constraints. Perhaps you are on a limited budget. Maybe you have to buy at least one present for each person on your list. You might have a lower limit on the money spent on a site (to get free shipping). You also can’t buy more of 11

an item than the merchant has. In this environment, you’re going to try to determine the number of items of each type that you buy from each merchant. The problem could become stochastic if there were a chance that a merchant might sell out of an item, or that deliveries are delayed, or that you may or may not need presents for certain people.

2. Politicians will often make statements like the following: “We are going to provide the best-possible health care at the lowest-possible cost.” While on its face this sounds like a laudable optimization problem, as stated this goal is actually nonsensical. Why? What would be a more accurate way to state the intended goal? Answer:

Lecture 1: The Operations Research Superhighway

It’s two goals. Assuming that we can’t have negative health-care costs, the lowest-possible cost is zero. But the best-possible health care is not going to cost zero. A more accurate way to state the goal would be to provide the best balance of health-care quality and cost. The trouble, of course, is that this immediately raises the question of who decides what that balance is, and how. This is exactly the kind of question that the politician might want not to address.

12

Forecasting with Simple Linear Regression Lecture 2

I

n this lecture, you will learn about linear regression, a forecasting technique with considerable power in describing connections between related quantities in many disciplines. Its underlying idea is easy to grasp and easy to communicate to others. The technique is important because it can—and does—yield useful results in an astounding number of applications. But it’s also worth understanding how it works, because if applied carelessly, linear regression can give you a crisp mathematical prediction that has nothing to do with reality. Making Predictions from Data z Beneath Yellowstone National Park in Wyoming is the largest active volcano on the continent. It is the reason that the park contains half of the world’s geothermal features and more than half of its geysers. The most famous of these is Old Faithful, which is not the biggest geyser, nor the most regular, but it is the biggest regular geyser in the park—or is it? There’s a popular belief that the ggeyser y erupts p once an hour,, like clockwork.

Figure 2.1

13

Lecture 2: Forecasting with Simple Linear Regression

z

In Figure 2.1, a dot plot tracks the rest time between one eruption and the next for a series of 112 eruptions. Each rest period is shown as one dot. Rests of the same length are stacked on top of one another. The plot tells us that the shortest rest time is just over 45 minutes, while the longest is almost 110 minutes. There seems to be a cluster of short rest times of about 55 minutes and another cluster of long rest times in the 92-minute region.

z

Based on the information we have so far, when tourists ask about the next eruption, the best that the park service can say is that it will probably be somewhere from 45 minutes to 2 hours after the last eruption—which isn’t very satisfactory. Can we use predictive modeling to do a better job of predicting Old Faithful’s next eruption WLPH":HPLJKWEHDEOHWRGRWKDWLIZHFRXOG¿QGVRPHWKLQJWKDW we already know that could be used to predict the rest periods.

z

$URXJKJXHVVZRXOGEHWKDWZDWHU¿OOVDFKDPEHULQWKHHDUWKDQG heats up. When it gets hot enough, it boils out to the surface, and then the geyser needs to rest while more water enters the chamber and is heated to boiling. If this model of a geyser is roughly right, we could imagine that a long eruption uses up more of the water in the chamber, DQGWKHQWKHQH[WUH¿OOUHKHDWHUXSWF\FOHZRXOGWDNHORQJHU:HFDQ make a scatterplot with eruption duration on the horizontal axis and the length of the following rest period on the vertical.

Figure 2.2

14

z

When you’re dealing with bivariate data (two variables) and they’re ERWKTXDQWLWDWLYHQXPHULFDO WKHQDVFDWWHUSORWLVXVXDOO\WKH¿UVW thing you’re going to want to look at. It’s a wonderful tool for exploratory data analysis.

z

Each eruption gets one dot, but that one dot tells you two things: the x-coordinate (the left and right position of the dot) tells you how long that eruption lasted, and the y-coordinate (the up and down position of the same dot) tells you the duration of the subsequent rest period.

z

We have short eruptions followed by short rests clustered in the lower left of the plot and a group of long eruptions followed by long rests in the upper right. There seems to be a relationship between eruption duration and the length of the subsequent rest. We can get a reasonable approximation to what we’re seeing in the plot by drawing a straight line that passes through the middle of the data,, as in Figure g 2.3.

Figure 2.3

z

7KLVOLQHLVFKRVHQDFFRUGLQJWRDVSHFL¿FPDWKHPDWLFDOSUHVFULSWLRQ :HZDQWWKHOLQHWREHDJRRG¿WWRWKHGDWDZHZDQWWRPLQLPL]H the distance of the dots from the line. We measure this distance vertically, and this distance tells us how much our prediction of rest time was off for each particular point. This is called the residual for that point. A residual is basically an error. 15

Lecture 2: Forecasting with Simple Linear Regression

Figure 2.4

16

z

7KH JUDSK KDV SRLQWV VR ZH FRXOG ¿QG WKHLU UHVLGXDOV² how well the line predicts each point. We want to combine these residuals into a single number that gives us a sense of how tightly the dots cluster around the line, to give us a sense of how well the line predicts all of the points.

z

You might think about averaging all of the distances between the dots and the line, but for the predictive work that we’re doing, it’s more useful to combine these error terms by squaring each residual before we average them together. The result is called the mean squared error (MSE). The idea is that each residual tells you how much of an error the line makes in predicting the height of a particular point—and then we’re going to square each of these errors, and then average those squares.

z

A small mean squared error means that the points are clustering tightly around the line, which in turn means that the line is a decent approximation to what the data is really doing. The straight line drawn in the Old Faithful scatterplot is the one that has the lowest MSE of any straight line you can possibly draw. The proper name for this prediction line is the regression line, or the least squares line.

z

Finding and using this line is called linear regression. More precisely, it’s simple linear regression. The “simple” means that we only have one input variable in our model. In this case, that’s the duration of the last eruption.

z

,I \RX NQRZ VRPH FDOFXOXV \RX FDQ XVH WKH GH¿QLWLRQ RI WKH regression line—the line that minimizes MSE—to work out the equation of the regression line, but the work is time consuming and tedious. Fortunately, any statistical software package or any decent spreadsheet, such as Excel RU 2SHQ2I¿FH¶V Calc FDQ ¿QG LW IRU you. In those spreadsheets, the easiest way to get it is to right-click on a point in your scatterplot and click on “add trendline.” For the cost of a few more clicks, it’ll tell you the equation of the line.

z

For the eruption data, the equation of the line is about y = 0.21x + 34.5, where x is the eruption duration and y is the subsequent rest. So, the equation says that if you want to know how long a rest to expect, on average, after an eruption, start with 34.5 minutes, and then add an extra 0.21 minutes for every additional second of eruption.

z

Any software package will also give you another useful number, the r2 value, which is also called the FRHI¿FLHQWRIGHWHUPLQDWLRQ, because it tells you how much the line determines, or explains, the data. For the Old Faithful data, the spreadsheet reports the r2 value as about 0.87. Roughly, that means that 87% of the variation in the height of the dots can be explained in terms of the line. In other words, the model explains 87% of the variation in rest times in terms of the length of the previous eruption.

Linear Regression z Linear regression assumes that your data is following a straight line, apart from “errors” that randomly bump a data point up or down from that line. If that model’s not close to true, then linear regression is going to give you nonsense. We’ll expect data to follow a straight line when a unit change in the input variable can be expected to cause a uniform change in the output variable. For Old Faithful, each additional second of eruption adds about 12 seconds of rest. 17

Lecture 2: Forecasting with Simple Linear Regression 18

z

If r2 is low, we’re on shaky ground—and that’s one thing everyone learns quite early about linear regression. But linear regression is so easy to do (at least with a statistical calculator or computer) that \RX¶OORIWHQVHHSHRSOHEHFRPLQJRYHUFRQ¿GHQWZLWKLWDQGJHWWLQJ themselves into trouble.

z

The problem is that linear regressions aren’t always as trustworthy as they seem. For example, using a small data set is a very bad way to make predictions. Even though you could draw a straight line between two data points and get an r2 of 1—a perfect straight-line ¿W²WKHOLQHWKDW\RX¿QGPLJKWEHDORQJZD\IURPWKHWUXHOLQH that you want, the one that gives the true underlying relationship between your two variables.

z

1RWRQO\PLJKWWKHOLQHWKDW\RX¿QGGLIIHUVLJQL¿FDQWO\IURPWKH true line, but the farther you get to the left or right of the middle of your data, the larger the gap between the true line and your line can be. This echoes the intuitive idea that the farther you are from your observed data, the less you can trust your prediction.

z

It’s a general principle of statistics that you get better answers from more data, and that principle applies to regression, too. But if so, how much data is enough? How much can we trust our DQVZHUV" $Q\ VRIWZDUH WKDW FDQ ¿QG WKH UHJUHVVLRQ HTXDWLRQ IRU you can probably also give you some insights into the answer to these questions. In Excel, it can be done by using the program’s regression report generator, part of its data analysis add-in. You put in your x and y values, and it generates an extensive report.

z

The software isn’t guaranteeing that the real intercept lies in the range it provides, but it’s making what is known as FRQ¿GHQFH interval predictions based on some often-reasonable assumptions about how the residuals are distributed. It’s giving a range that is 95% likely to contain the real intercept.

z

The uncertainties in the slope and intercept translate into uncertainties in what the correct line would predict. And any LQDFFXUDF\RIWKHOLQHJHWVPDJQL¿HGDVZHPRYHIDUWKHUIURPWKH center of our data. The calculations for this are a bit messy, but if your data set is large and you don’t go too far from the majority of your sample, the divergence isn’t going to be too much.

z

6XSSRVHWKDWZHZDQWWREHFRQ¿GHQWDERXWWKHYDOXHRIRQH variable, given only the value of the second variable. There’s a complicated formula for this prediction interval, but if your data set is large, there’s a rule of thumb that will give you quite a good working approximation. Find one number in your regression report: It’s usually called either the standard error or standard error of the regression. Take that number and double it. About 95% of the time, the value of a randomly selected point is going to be within this number’s range of what the regression line said.

z

So, if you’re talking about what happens on average, the regression line is what you want. If you’re talking about an individual case, you want this prediction interval.

Important Terms Calc7KH2SHQ2I¿FHVXLWH¶VHTXLYDOHQWWR([FHO,W¶VIUHHO\GRZQORDGDEOH but lacks some of the features of Excel. cluster: A collection of points considered together because of their proximity to one another. FRHI¿FLHQWRIGHWHUPLQDWLRQ: See r2. FRQ¿GHQFH LQWHUYDO: An interval of values generated from a sample that hopefully contains the actual value of the population parameter of interest. See FRQ¿GHQFHVWDWLVWLFV .

19

error: In a forecasting model, the component of the model that captures the variation in output value not captured by the rest of the model. For regression, this means the difference between the actual output value and the value forecast by the true regression line. Excel7KH0LFURVRIW2I¿FHVXLWH¶VVSUHDGVKHHWSURJUDP linear regression$PHWKRGRI¿QGLQJWKHEHVWOLQHDUUHODWLRQVKLSEHWZHHQD set of input variables and a single continuous output variable. If there is only one input variable, the technique is called simple; with more than one, it is called multiple.

Lecture 2: Forecasting with Simple Linear Regression

prediction interval7KH SUHGLFWLRQ LQWHUYDO LV DQ LQWHUYDO ZLWK D VSHFL¿HG probability of containing the value of the output variable that will be REVHUYHGJLYHQDVSHFL¿HGVHWRILQSXWV&RPSDUHWRFRQ¿GHQFHLQWHUYDO. r2 7KH FRHI¿FLHQW RI GHWHUPLQDWLRQ D PHDVXUH RI KRZ ZHOO D IRUHFDVWLQJ model explains the variation in the output variable in terms of the model’s inputs. Intuitively, it reports what fraction of the total variation in the output variable is explained by the model. regression: A mathematical technique that posits the form of a function FRQQHFWLQJ LQSXWV WR RXWSXWV DQG WKHQ HVWLPDWHV WKH FRHI¿FLHQWV RI WKDW function from data. The regression is linear if the hypothesized relation is linear, polynomial if the hypothesized relation is polynomial, etc. regression line: The true regression line is the linear relationship posited to exist between the values of the input variables and the mean value of the output variable for that set of inputs. The estimated regression line is the approximation to this line found by considering only the points in the available sample. residual: Given a data point in a forecasting problem, the amount by which the actual output for that data point exceeds its predicted value. Compare to error.

20

sample: A subset of a population. standard error: Not an “error” in the traditional sense. The standard error is the estimated value of the standard deviation of a statistic. For example, the standard error of the mean for samples of size 50 would be found by JHQHUDWLQJ HYHU\ VDPSOH RI VL]H IURP WKH SRSXODWLRQ ¿QGLQJ WKH PHDQ of each sample, and then computing the standard deviation of all of those sample means.

Suggested Reading Hyndman and Athanasopoulos, Forecasting. Miller and Hayden, Statistical Analysis with the General Linear Model. Ragsdale, Spreadsheet Modeling & Decision Analysis.

Questions and Comments 1. Imagine that we set a group of students on a task, such as throwing 20 darts and trying to hit a target. We let them try, record their number of successes, and then let them try again. When we record their results in a scatterplot, we are quite likely to get something similar to the following graph. The slope of the line is less than 1, the students who did the best RQWKH¿UVWWU\WHQGWRGRZRUVHRQWKHVHFRQGDQGWKHVWXGHQWVZKRGLG ZRUVWRQWKH¿UVWWU\WHQGWRLPSURYHRQWKHVHFRQG,IZHSUDLVHGWKH VWXGHQWVZKRGLGZHOORQWKH¿UVWWU\DQGSXQLVKHGWKRVHZKRGLGSRRUO\ we might take these results as evidence that punishment works and praise is counterproductive. In fact, it is just an example of regression toward the mean. (See Figure 2.5.) Assume that a student’s performance is a combination of a skill factor and a luck factor and that the skill factor for a student is unchanged from trial to trial. Explain why you would expect behavior like that suggested by the graph without any effects of punishment or praise.

21

Lecture 2: Forecasting with Simple Linear Regression

Figure 2.5

Answer: Consider the highest scorers in the original round. Their excellence is probably due to the happy coincidence of considerable skill and considerable luck. When this student repeats the exercise, we can expect the skill factor to be essentially unchanged, but the luck factor is quite OLNHO\WRGHFUHDVHIURPWKHXQXVXDOO\KLJKYDOXHLWKDGLQWKH¿UVWURXQG The result is that the performance of those best in round 1 is likely to decrease in round 2. On the low end, we have a mirror of this situation. The worst performers probably couple low skill with bad luck in round 1. That rotten luck is likely to improve in round 2—it can hardly get worse! This effect is seen in a lot of real-life data. For example, the children of the tallest parents are usually shorter than their parents, while the children of the shortest parents are usually taller than their parents.

22

2. Suppose that you are given a sack that you know contains 19 black marbles and 1 white marble of identical size. You reach into the bag, close your hand around a marble, and withdraw it from the bag. It is FRUUHFWWRVD\WKDW\RXDUHFRQ¿GHQWWKDWWKHPDUEOHLQ\RXUKDQG LV EODFN DQG LW LV LQ WKLV VHQVH WKDW WKH WHUP ³FRQ¿GHQFH´ LV XVHG LQ VWDWLVWLFV&RQVLGHUHDFKRIWKHVWDWHPHQWVEHORZDQG¿QGWKHRQHWKDWLV HTXLYDOHQWWR\RXU³FRQ¿GHQFH´VWDWHPHQW a)

This particular marble is 95% black and 5% white. (Maybe it has white spots!)

b) This particular marble is black 95% of the time and white 5% of the WLPH3HUKDSVLWÀLFNHUV c)

This particular marble doesn’t have a single color, only a probability. Its probability of being black is 95%.

d) The process by which I got this particular marble can be repeated. If it were repeated many, many times, the resulting marble would be black in about 95% of those trials. Answer: The answer is d), but the point of the question is that answers a) through c) correspond roughly to statements that are often made by people when LQWHUSUHWLQJFRQ¿GHQFH)RUH[DPSOHJLYHQDFRQ¿GHQFHLQWHUYDO for mean income as $40,000 to $50,000, people will often think that 95% of the population makes money between these bounds. Others will say that the mean is in this range 95% of the time. (The mean of the SRSXODWLRQLVDVLQJOH¿[HGQXPEHUVRLWLVHLWKHULQWKHLQWHUYDORULW LV QRW :KHQ ZH GHFODUH FRQ¿GHQFH ZH DUH VSHDNLQJ RI FRQ¿GHQFH in a process giving an interval that manages to capture the population parameter of interest.

23

Nonlinear Trends and Multiple Regression Lecture 3

T

Lecture 3: Nonlinear Trends and Multiple Regression

here are two important limitations to simple linear regression, both of which will be addressed in this lecture. First, linear regression is fussy about the kind of relation that connects the two variables. It has to be linear, with the output values bumped up and down from that straight-line relation by random amounts. For many practical problems, the scatterplot of input versus output looks nothing like a straight line. The second problem is that simple linear regression ties together one input with the output. In many situations, the values of multiple input variables are relevant to the value of the output. As you will learn, multiple linear regression allows for multiple inputs. Once these tools are in place, you can apply them to nonlinear dependencies on multiple inputs. Exponential Growth and Decay z Exponential growth is going to show up any time that the rate at which something is growing is proportional to the amount of that something present. For H[DPSOH LQ ¿QDQFH LI \RX KDYH twice as much money in the bank at the beginning of the year, you earn twice as much interest during that year. Exponential decay shows up when the rate at which something is shrinking is proportional to the amount of that something present. For example, in advertising, if there are only half as many customers left to reach, your ads are only reaching half as many new customers.

24

Figure 3.1

Figure 3.2

z

For exponential growth, the time taken for the quantity to double is a constant. For example, Moore’s law, which states that the number of transistors on a microchip doubles every two years, describes exponential growth. For exponential decay, the amount of time required for something to be cut in half is constant. For example, half-life for radioactivity is exponential decay.

z

Anything undergoing exponential growth or decay can be expressed mathematically as y = cax + b, where y is the output (the quantity that’s growing or shrinking); x is the input (in many models, that’s time); and a, b, and c are constants. You can pick a value for c; anything bigger than 1 is a good workable choice.

z

So many things follow the kind of hockey-stick curve that we see in exponential growth or decay that we really want to be able to predict them. Unfortunately, at the moment, our only prediction technique is restricted to things that graph as straight lines: linear expressions. In algebra, y = ax + b.

z

Anytime you do algebra and want to solve for a variable, you always have to use inverse functions—functions that undo what you’re trying to get rid of. You can undo an exponentiation by using its inverse: logarithm (log). If you take the log base c of both sides, logc y = logc (cax + b ZKLFKVLPSOL¿HVWRORJc y = ax + b. This results in a linear expression on the right side of the equation, but y is no longer on the left—instead it’s the log of y.

z

If y is a number that we know and c is a number that we know, then the logc yLVMXVWDQXPEHUWRR²RQHZHFDQ¿QGZLWKDVSUHDGVKHHW or calculator using a bunch of values for x and y. Whereas x versus y will graph as an exponential, x versus log y will graph as a straight line. And that means that if you start with x and y values that are close to an exponential relationship, then x and log y will have close to a linear relationship—and that means that we can use simple linear regression to explore that relationship.

25

Lecture 3: Nonlinear Trends and Multiple Regression

z

This works for any reasonable c that you pick—anything bigger than 1 will work, for example. Most people use a base that is a number called e: 2.71828…. Using this base makes a lot of more advanced work a lot easier.

z

No matter what base we use, we’re going to need a calculator or VSUHDGVKHHW WR ¿QG SRZHUV DQG ORJDULWKPV DQG FDOFXODWRUV DQG spreadsheets have keys for e. Most calculators have an ex key, along with a key for the log base e, which is also called the natural logarithm (ln). The loge x, the natural log of x, or the ln x all mean the same thing. And ln and e to a power are inverses—they undo one another.

Power Laws z Exponential growth and decay are a family of nonlinear relationships that can be analyzed with linear regression by a simple transformation of the output variable—by taking its logarithm. But there’s another family of relationships that are perhaps even more common that will yield to an extended application of this same idea. z

Suppose that we took the log of both the input and output variables. We’d be able to apply linear regression to the result if ln x and ln y actually do have a linear relationship. That is, if ln y = a ln x + b, where a and b are constants. Then, using laws of exponents and the fact that e to the x undoes ln, we can recover the original relation between x and y, as follows. ln y = a ln x + b eln y = ea ln x + b = ea ln xeb y = eb(eln x)a = ebxa

z

26

Therefore, the relationship between y and x is y = ebxa, and eb is just a positive constant, so we’re saying that y is proportional to VRPH¿[HGSRZHURIx. A relationship where one variable is directly proportional to a power of another is called a power law, and such UHODWLRQVKLSV DUH UHPDUNDEO\ FRPPRQ LQ VXFK ¿HOGV DV VRFLRORJ\

neuroscience, linguistics, physics, computer science, geophysics, economics, and biology. You can discover whether a power law is a decent description of your data by taking the logarithm of both variables and plotting the results. z

So many relationships seem to follow a rough power relation that research is being done as to why these kinds of connections should appear so often. But whenever they do, a log-log plot can tip you RIIWRLWDQGOLQHDUUHJUHVVLRQFDQOHW\RX¿QGWKHHTXDWLRQWKDW¿WV

Multiple Regression z What about allowing more than one input? With a linear relationship, each additional input variable adds one dimension of space to the picture, so the “best straight line through the data” picture needs to change, but the idea of linear regression will remain the same. The mathematics of this plays the same game that we used for simple linear regression. z

Actually doing the math for this becomes quite tedious. The good news is that, again, statistical software or spreadsheets can do the work for you easily. If you’re using a spreadsheet, Excel’s report has historically been more complete and easier to read than 2SHQ2I¿FH&DOF¶VEXWERWKFDQGRWKHMRE$QGVWDWLVWLFDOVRIWZDUH like R—which is free online—can do an even more thorough job.

z

It’s important to note that the FRHI¿FLHQW of a variable in a model is intended to capture the effect of that variable if all other inputs DUHKHOG¿[HG7KDW¶VZK\ZKHQWZRYDULDEOHVPHDVXUHDOPRVWWKH same thing, it’s often a good idea not to include both in your model. Which one gets credit for the effect can be an issue. This is a special case of the problem of multicollinearity.

z

Another variant of linear regression is called polynomial regression. Suppose that you have bivariate data that suggests a nonlinear relationship from the scatterplot and that your “take the log” transformations can’t tame into a straight line. Multiple

27

UHJUHVVLRQJLYHV\RXDZD\RI¿WWLQJDSRO\QRPLDOWRWKHGDWD7KHUH is a lot going on in multiple regression, and there is some pretty sophisticated math that supports it.

Important Terms FRHI¿FLHQW7KHQXPEHUPXOWLSOLHGE\DYDULDEOHLVLWVFRHI¿FLHQW e: A natural constant, approximately 2.71828. Like the more familiar ʌ, e appears frequently in many branches of mathematics.

Lecture 3: Nonlinear Trends and Multiple Regression

exponential growth/decay: Mathematically, a relationship of the form y = abx for appropriate constants a and b. Such relations hold when the rate of change of a quantity is proportional to its current value. linear expression: An algebraic expression consisting of the sum or difference of a collection of terms, each of which is either simply a number RUDQXPEHUWLPHVDYDULDEOH/LQHDUH[SUHVVLRQVJUDSKDV³ÀDW´REMHFWV² straight lines, planes, or higher-dimensional analogs called hyperplanes. logarithm: The inverse function to an exponential. If y = ax for some positive constant a, then x = loga y. The most common choice for a is the natural constant e. loge x is also written ln x. multicollinearity: The problem in multiple regression arising when two or more input variables are highly correlated, leading to unreliable estimation RIWKHPRGHOFRHI¿FLHQWV polynomial: A mathematical expression that consists of the sum of one or more terms, each of which consists of a constant times a series of variables raised to powers. The power of each variable in each term must be a nonnegative integer. Thus, 3x2 + 2xy + zíLVDSRO\QRPLDO power law: A relationship between variables x and y of the form y = axb for appropriate constants a and b.

28

Suggested Reading Hyndman and Athanasopoulos, Forecasting. Miller and Hayden, Statistical Analysis with the General Linear Model.

Questions and Comments 1. 7KH OHFWXUH PHQWLRQHG WKDW RQH FRXOG XVH OLQHDU UHJUHVVLRQ WR ¿W D polynomial to a set of data. Here, we look at it in a bit more detail. Given a table of values for the input x and the output y, add new input variables whose values are x2, x3, and so on. Stop when you reach the degree of polynomial that you wish to use. Now conduct multiple regression in the normal way with these variables. The table used in the regression might begin as follows. x

x2

x3

x4

y

3

9

27

81

17

í

4

í

16

40

…

…

…

…

… Figure 3.3

The same technique can be used to look for interaction effects between two different input variables. In addition to input variables x1 and x2, for example, we could include the interaction term x1x2. For example, LQFOXGLQJ HLWKHU PXVWDUG RU -HOO2 LQ D GLVK PLJKW HDFK EH ¿QH individually but might create quite an unpleasant reaction together!

2. ,Q PRVW RI LWV LQFDUQDWLRQV UHJUHVVLRQ LV SUHWW\ VSHFL¿F DERXW ZKDW the “random errors” in a model are supposed to look like. You could imagine how they’re supposed to work in this way. Suppose that you have a bucket containing a huge number of poker chips, each with a number on it. The numbers are centered on zero, balanced out between positive and negative, and there are a lot more chips with values close to 29

zero than there are with values of large magnitude. When you need the error for a particular input point, reach into the bucket for a chip, read its number, and then add that number to the calculated linear output. Then, throw the poker chip back in the bucket.

Lecture 3: Nonlinear Trends and Multiple Regression

More technically, the errors are supposed to be normally distributed with a mean of zero, a constant standard deviation, and are supposed to be uncorrelated to one another as well as being uncorrelated to the input values—but the error bucket gets the key idea across.

30

Time Series Forecasting Lecture 4

T

he topic of this lecture is forecasting—predicting what’s going to happen, based on what we know. In many circumstances, we’re looking at historical data gathered over time, with one observation IRUHDFKSRLQWLQWLPH2XUJRDOLVWRXVHWKLVGDWDWR¿JXUHRXWZKDW¶VJRLQJ to happen next, as well as we can. Data of this type is called time series data, and to have any hope of making progress with predicting time series data, we have to assume that what has gone on in the past is a decent model for what will happen in the future. Time Series Analysis z Let’s look at some historical data on U.S. housing starts—a monthby-month record of how many new homes had their construction start in each month. Housing starts are generally considered to be a leading indicator of the economy as a whole. z

For a time series, we can visualize the data by making a line graph. The horizontal axis is time, and we connect the dots, where each dot represents the U.S. housing starts for that month. The basic strategy is to decompose the time series into a collection of different components. Each component will capture one aspect of the historical behavior of the series—one part of the pattern.

Figure 4.1

31

Lecture 4: Time Series Forecasting 32

z

The variation in the data series—the up-and-down bouncing—is far from random. Each January, new housing starts tank, then climb rapidly in the spring months, reaching a peak in summer. Given the weather patterns in North America, this makes sense, and we’d have every reason to expect this kind of variation to continue into the future.

z

:H¶YH MXVW LGHQWL¿HG WKH ¿UVW FRPSRQHQW RI RXU WLPH VHULHV decomposition: the seasonal component. Seasonal components are SDWWHUQVWKDWUHSHDWRYHUDQGRYHUDOZD\VZLWKD¿[HGGXUDWLRQMXVW like the four seasons. But the period of repetition doesn’t have to be D\HDULWFDQEHDQ\UHJXODUYDULDWLRQRI¿[HGGXUDWLRQ

z

Getting a handle on seasonality is important in two ways. First, if you’re hoping to make accurate forecasts of what’s going to happen at some point in the future, then you’d better include seasonal variation in that forecast. Second, when trying to make sense of the past, we GRQ¶W ZDQW VHDVRQDO ÀXFWXDWLRQV WR FRQFHDO RWKHU PRUHSHUVLVWHQW trends. This is certainly the case with housing starts and why the government reports “seasonally adjusted” measures of growth.

z

The other obvious pattern in the data, once seasonality is accounted for, is that there appears to be a steady increase in housing starts. In fact, we can apply simple linear regression to this line to see how ZHOODOLQHDUWUHQG¿WVWKHGDWD,QWKLVH[DPSOHx is measured in months, with x = 1 being January 1990, x = 13 being January 1991, and so on.

z

With r2 being only 0.36, about 36% in the variation in housing starts can be laid at the doorstep of the steady passage of time. That leaves 64% unaccounted for. But this is what we expect. The data has a very strong annual seasonal component, and the trend line is going to completely ignore seasonal effects. In the sense of tracking the center of the data, the regression line actually seems to be doing rather well.

Figure 4.2

z

For this example, the regression line would be the second component of the time series, the trend component. Not all data demonstrates a linear trend, and in general, trend components can actually be quite complicated.

z

There’s a third component that arises with some time series called the cyclic component, and it tracks cyclic variation. While cyclic variation can be thought of as including regular seasonality as one of its subtypes, it’s clearer to say that cyclic variation refers WR ORQJHUWHUP ÀXFWXDWLRQV WKDW ODFN WKH UHJXODULW\ RI VHDVRQDO variation. Business cycles are a good example: growth, recession, recovery—but of variable onset, intensity, and duration.

z

Our data for housing starts doesn’t show any cyclic variation. In fact, many short- and medium-range techniques for forecasting two years or fewer don’t include a cyclic component. So, we’re currently modeling our housing starts as a seasonal component overlaid on a linear trend.

z

Just like in regression, a time series almost never perfectly matches the real-world data. Whatever variation is left unexplained is LGHQWL¿HG DV WKH HUURU FRPSRQHQW ZKLFK HVVHQWLDOO\ FRQVLVWV RI residuals. The component captures all of the variation between what the model predicts and what actually happens.

33

Lecture 4: Time Series Forecasting

z

When we do a good job with our forecasting, there shouldn’t be DQ\ VLJQL¿FDQW SDWWHUQ WR WKH HUURU FRPSRQHQW 3DWWHUQV LQ WKH error component mean that the errors contain more information— information that we could have squeezed out of them and included in the other components of the forecast.

z

Everything described so far applies to almost every time series forecasting technique, but there are a large number of such techniques, and the variety exists for a reason. Different time series display different characteristics, and the processes that generate them may dictate restrictions on the kind of model that we use.

z

Not every model includes all of the possible kinds of components. Some data shows virtually no seasonality, and some shows virtually no trend. But suppose that you have data that includes both. How do you combine them?

z

Two common approaches are additive and multiplicative models. In an additive model, you say that the observed value is the sum of the trend component, the seasonal component, and the error component. This is good when the seasonal component stays pretty constant over time. In a multiplicative model, you multiply these pieces together instead of adding them—a better choice when the seasonal component’s magnitude varies with the trend.

Measures of Forecast Quality z People who do forecasting like to be a bit more quantitative when assessing the performance of a forecast. One common measure is called PHDQDEVROXWHGHYLDWLRQ0$' 6WDUWE\¿QGLQJKRZPXFK each forecast value differed from the actual historical value—the error for each point in time. If it’s negative, take its absolute value to make it positive. Finally, average all of these absolute errors.

34

z

In addition to the MAD, people also often report the mean absolute SHUFHQWDJH HUURU 0$3( 7R GR WKLV ¿QG WKH SHUFHQW WKDW HDFK forecast was wrong, take the absolute value of each of these in case they’re negative, and then average all of these percentages. MAD and MAPE are both particularly popular in the manufacturing sector.

z

%XW WKH PRVW FRPPRQ ZD\ WR FKDUDFWHUL]H WKH ¿W RI D IRUHFDVW LV PHDQ VTXDUHG HUURU 06( . That is, take the error for each observation, square each of these errors, and then average all of the squared errors together.

z

MSE has much nicer statistical and calculus properties than MAD or MAPE does. It falls short, though, when you try to interpret it simply. You can’t compare these different measures of forecast quality to one another, but you can compare two different forecasts using the same measure of quality. The one with the lower MAD— or MAPE, or MSE—should be the better forecast.

z

Often, all of the different measures agree on the winner, but not always. MSE tends to care about consistency in a forecast. If you’re exactly right a lot of the time but occasionally make howling errors, MSE gives you a big penalty for the mistakes. That is, its squaring step turns a very bad error into a very, very bad error. MAD and MAPE are more forgiving of the occasional howler. In many applications, people would prefer many small errors to an occasional large one, which is one of the non-computational reasons that MSE is often the preferred measure.

Characterizing Trends in Data z One of the dangers you face when you’re working with historical data DORQHLVWKDWLWODFNVDVWUXFWXUDOPRGHORIZKDWIDFWRUVDUHLQÀXHQFLQJ the evolution of the data outputs. Time series forecasting is based on the idea that the past is a good model for the future. When something fundamental about the situation changes, if you can’t anticipate it, you can be left with forecasts that are really quite dreadful.

35

Lecture 4: Time Series Forecasting 36

z

Extending a graph over a much longer number of years also shows how the business cycle (or other cyclic variation) can turn out to be important, even for short-term forecasts. One reason the business F\FOHLVRIWHQLJQRUHGLVWKDWHFRQRPLVWV¿QGLWVRGLI¿FXOWWRVSHFLI\ precisely. The better we understand the past, the better chance we have of predicting the future. And, sometimes, such as in economic analysis, we’re also trying to make sense of the past for its own sake.

z

A common way of characterizing a trend in a group of data is by using a simple moving average. Simply put, we peg the trend at some point in time by averaging together some number of observations near that point in time. If you’re trying to forecast the future with monthly data, you might average together the 12 most recent months to get the forecast of what happens next. If you’re trying to make sense of historical data, you might choose a set of observations centered on your current month to average.

z

But because each simple moving average forecast is just the average of the preceding 12 months, we don’t get a forecast until 12 months have gone by. More importantly, we can only use the technique to forecast one month in advance. If we want to go farther than that— and we probably do—we need more advanced techniques.

z

And a plethora of them exist, each suited to different kinds of time series. Some are simple. Weighted moving average takes the simple moving average but gives each observation a different weight. These weights tell us the relative importance of the values used in computing the forecast. If what happened one time period DJRKDVWZLFHDVPXFKLQÀXHQFHRQWKHSUHVHQWDVZKDWKDSSHQHG WZRWLPHSHULRGVDJRWKHQWKHLUZHLJKWVZRXOGUHÀHFWWKLV

z

2IFRXUVH\RXKDYHWR¿QGWKHZHLJKWVWKDWPDNHWKHIRUHFDVW¿W WKHKLVWRULFDOGDWDDVZHOODVSRVVLEOH7KDWLV\RX¶GOLNHWR¿QGWKH values of the weights that minimize some measurement of error, such as the MSE of the forecast.

z

A close relative of the weighted moving average is called simple exponential smoothing. You can think of it as a weighted moving average in which weights grow smaller and smaller in a geometric fashion as we move back in time. Exponential smoothing is an extremely simple forecasting technique, but it’s the basis for a lot of more sophisticated and complicated approaches.

Important Terms cyclic component: The component of a time series forecast that attempts to capture cyclic variation. This differs from seasonal variation by showing nonconstant duration or intensity or unpredictable onset. exponential smoothing: A time series forecasting technique that forms the basis of many more-complicated models. Exponential smoothing can be thought of as a variant of the weighted moving average. PHDQDEVROXWHGHYLDWLRQ0$' : A measure of forecast accuracy, MAD is the average amount by which the forecast differs from the actual value. PHDQDEVROXWHSHUFHQWDJHHUURU0$3( : A measure of forecast accuracy, MAPE is the average percentage by which the forecast differs from the actual value. PHDQVTXDUHGHUURU06( : A measure of forecast accuracy that is similar to variance in its calculation, MSE is the average of the squares of all of the residuals for a forecast. penalty $ PRGL¿HU WR WKH REMHFWLYH RI D SUREOHP WKDW UHÀHFWV WKDW WKH SURSRVHGVROXWLRQLVGH¿FLHQWLQFRPSOHWHO\VDWLVI\LQJDFRQVWUDLQW8VHGLQ genetic algorithms and soft constraints. seasonal component: The component of a time series forecast that captures the seasonality of the data—that is, its regular, periodic variation. Some sources use the term for such variation only if the period length is at least one year.

37

simple moving average: A forecast for a period in a time series made by DYHUDJLQJWRJHWKHUWKHYDOXHVIRUDVSHFL¿HGQXPEHURIQHDUE\WLPHSHULRGV For predictive models, this will mean the n observations immediately preceding the current time period. time series: A data set consisting of one value of the output variable for each point in time. The points in time are usually evenly spaced. Alternatively, a forecasting technique used on such data. weighted moving average: A forecast for a period in a time series made by DYHUDJLQJWRJHWKHUWKHYDOXHVIRUDVSHFL¿HGQXPEHURIQHDUE\WLPHSHULRGV ZLWKHDFKSHULRG¶VLPSRUWDQFHUHÀHFWHGE\LWVZHLJKW)RUSUHGLFWLYHPRGHOV this will mean a weighted average of the n observations immediately preceding the current time period.

Suggested Reading Hyndman and Athanasopoulos, Forecasting.

Questions and Comments

Lecture 4: Time Series Forecasting

1. If you’re going to apply more complicated time series approaches, you’re probably going to want to use software to do it. A few resources that you PLJKW¿QGH[WUHPHO\KHOSIXOLQOHDUQLQJPRUHRQ\RXURZQLQFOXGH the open-source statistical software package R, which will run on a wide variety of computer platforms and is fairly straightforward to learn, powerful, and free; and 2) Forecasting: Principles and Practice, a free online text that does an excellent job of explaining more sophisticated time series analysis in more detail, including the R commands necessary to conduct the analysis on a time series.

2. It is possible to look at two different time series over the same interval of time and to explore their relationship in a scatterplot. The point (xi, yi) is plotted if at time period iWKH¿UVWVHULHVKDVWKHYDOXHxI and the second series has the value yi.

38

Think about what would happen if x and y really had no relationship to one another but both showed a positive trend over a period of time. The resulting scatterplot could have a high correlation for the two variables, which could mislead the investigator into believing the two variables were in fact linked. For this reason, it’s dangerous to create such a scatterplot for time series with trends.

39

Data Mining—Exploration and Prediction Lecture 5

S

Lecture 5: Data Mining—Exploration and Prediction

tatistical techniques like regression and time series forecasting were strongholds of mathematical prediction for much of the 20th century. But they have been supplemented in recent decades by additional techniques from the exciting and fast-growing area known as data mining, which focuses on large data sets. As you will learn in this lecture, the job RI GDWD PLQLQJ LV WR ¿QG ZKDW XVHIXO SDWWHUQV DQG FRQQHFWLRQV ODUJH GDWD sets contain—connections and patterns that might otherwise be missed. Data mining is especially important in the 21st century because of two advances: computational power and a veritable explosion in the quantity of collected data. Data Mining z One of the key differences between many classical statistical techniques and data mining is the quantity of data available. Throughout much of history, data was scarce and hard to come by. And that meant that all of the data that was available had to be used in the analysis—both to create the model and to test its accuracy.

40

z

In order for that analysis to get very far, some assumptions had to be made, including that the errors are normally distributed (in technical terms, the data is supposed to be homoscedastic) and that the errors are independent from one another. The violation of these assumptions leads to questions of how much to trust the predictions.

z

In data mining, we often have an embarrassment of riches when LW FRPHV WR GDWD²LI DQ\WKLQJ ZH KDYH WRR PXFK 7KH ¿UVW WKLQJ that means is that we can use part of the data to come up with our model and a completely different set of data to test the reliability of that model. This also means that we aren’t bound by a lot of assumptions, such as the distribution of errors.

z

In other techniques, such as regression, we can spend a lot of time transforming the data so that we don’t have heteroscedasticity, autocorrelation, or error terms that are distributed, for example, like lottery payoffs—many small negative values sprinkled with occasional big positive values. Fixing these problems can be quite a headache. With data mining, the headache largely goes away.

z

Practitioners sometime debate where exactly the margins of data mining lie, but there are three objectives that are central to it: data visualization and exploration, FODVVL¿FDWLRQ and prediction, and association among cases. Time series analysis is sometimes considered a part of data mining, too.

Data Visualization z When we’re exploring data in data mining, we’re probably going to start by looking at one variable at a time. If a variable is categorical, bar charts are a good choice. In the following chart, we can see at a glance that heart disease and cancer—“malignant neoplasms”—are overwhelmingly the most common causes of death in the United States, and seeing this in a chart conveys this fact much more quickly and memorably than a table of numbers does.

Figure 5.1

41

z

Increasingly, software supports the ability to look at interesting aspects of the data in greater detail. For example, from the previous graph, we can break things down by age, as follows. Excel’s pivot table and pivot chart can give you some of this functionality.

Lecture 5: Data Mining—Exploration and Prediction

Figure 5.2

z

For a single numeric variable, the most common choice is the histogram, which is like a bar chart in which each bar represents a range of values. For example, the following is a histogram of the distribution for how many Facebook friends a group of students have.

Figure 5.3

42

z

For relationships between two numeric variables, we generally use scatterplots. This initial exploration often uncovers connections between the variables—connections we can exploit in further analysis.

z

For relationships among three variables, we can use a threedimensional plot. This is less useful when the software renders such data as a static plot on a two-dimensional screen, or piece of paper. But there are still options. We can code additional information with color or size. Many weather maps do this. Such representations are often called heat maps, even when the color is coding for something quite different than temperature or when the thing being colored is not a map.

z

We can also use the time dimension, combining multiple images into a movie. Swedish academician Hans Rosling created an DVWRXQGLQJ YLGHR WKDW \RX FDQ ¿QG RQ WKH %%& )RXU ZHEVLWH LQ which he tracks the wealth and longevity of people in 200 countries over 200 years of history. He uses a scatterplot with an axis for lifespan, a logarithmic axis for wealth, color to indicate continent, size to indicate population, and time to indicate the passage of \HDUV7KH¿OPODVWVRQO\IRXUPLQXWHVDQGLVVWXQQLQJO\HIIHFWLYHDW communicating all of that data.

z

Processing visual information in two or three dimensions— recognizing structures and patterns—is something that human beings are uncannily good at, which is why charts and graphs exist to begin with. But as the number of variables increases, obvious ways of representing data graphically fail. Each variable—each category of information—needs its own dimension in space, and for visualization, three dimensions of physical space is our limit. So, what can we do?

z

To begin, we can conduct a kind of dimensional reduction by using mathematical procedures. It’s like using a student’s GPA or SAT score in place of every detail of his or her academic history. The WULFNLVWR¿QGDZD\WRVXPPDUL]HPXOWLSOHSLHFHVRIGDWDLQDZD\ that loses as little information as possible. 43

z

The computations required may not be simple, but because data PLQLQJ E\ GH¿QLWLRQ GHDOV ZLWK ODUJH GDWD VHWV QR RQH ZRXOG perform the calculations by hand anyway. Statistical software such as SAS or SPSS, or the free package known as R, are perfectly capable of this kind of dimensional reduction; in fact, it can even be done in Excel with an add-in package such as XLMiner.

Lecture 5: Data Mining—Exploration and Prediction

Measuring Variation and Normalizing Data z We measure variation by using the most common statistical measure of dispersion: the variance. The calculation for the variance bears a great similarity to the calculation of MSE—that “average the squared errors” calculation for regression. For variance, the only difference is that the “error” is taken to be the distance of each data point from the mean of the data. If you take the square root of the variance, you get the standard deviation, something that comes up in most statistical discussions.

44

z

A technique called SULQFLSDOFRPSRQHQWVDQDO\VLV3&$ creates new variables that partition the variation between two variables LQ D EHWWHU ZD\ ZLWK WKH ¿UVW YDULDEOH FDSWXULQJ DV PXFK RI WKH YDULDWLRQDVSRVVLEOH3&$EHJLQVE\¿QGLQJWKHOLQHDORQJZKLFK that variation—that variance—is maximized. The math can be done by a computer, which gives us a line that would normally be called the z1 axis when doing PCA. Impressively, this line captures over 97% of the total variance in the two variables. To capture the remaining 2.6%, PCA would construct a second axis, perpendicular WRWKH¿UVWFDOOHGWKHz2 axis.

z

In complicated examples, with more than a handful of variables, 3&$ ¿QGV WKH GLUHFWLRQ LQ PXOWLGLPHQVLRQDO VSDFH WKDW FDSWXUHV the largest possible fraction of the variation in the data and calls that the z1 axis. It then looks for a direction perpendicular to that one that captures the maximal amount of the remaining variation. That’s the z2 axis, and with more than two dimensions in the original problem, there are more perpendicular directions to choose

DPRQJ3&$FRQWLQXHVOLNHWKLV,WQH[W¿QGVDGLUHFWLRQIRUDz3 axis that is perpendicular to both z1 and z2 and captures as much of the remaining variation as possible—and so on. z

:HRIWHQ¿QGWKDWDIWHURQO\DKDQGIXORISULQFLSDOFRPSRQHQWVDUH found, the vast majority of the variation is accounted for. If that’s the case, the original variables, however many there were, can be discarded and replaced with this handful, losing almost nothing about the original variation in the data and greatly simplifying the future work.

z

But there are some disadvantages to doing this, too. For example, PCA won’t work well for data whose underlying pattern is not linear. But that can be overcome by modeling a nonlinear pattern using local patterns that are linear, which is analogous to multiple regression.

z

Normalizing GDWD PHDQV VWDQGDUGL]LQJ LW :H GR WKLV E\ ¿UVW subtracting the mean of a variable from all values of that variable, and then dividing each of those differences by the standard deviation of that variable. No matter what data you start with, after you’ve normalized it, the mean of the data is zero, and its standard deviation is one. In that sense, then, each variable is now on equal footing. Generally, if the two variables are measured in units that aren’t comparable or on dramatically different scales, normalization is a good idea.

z

Another factor that can be useful in reducing the dimension of WKHGDWDLVGRPDLQNQRZOHGJH6RPHRQHIDPLOLDUZLWKWKH¿HOGRI investigation can often narrow our focus before we even start with the math. Experts can tell us what quantities are almost certainly relevant—or irrelevant—to the task at hand. They can also let us know when one of our variables is useless as a predictor because we won’t be able to gather it in time to make use of it. Domain NQRZOHGJHFDQDOVRDOORZXVWRLQVSHFWRXU¿QDODQVZHUDQGVHHLI it makes sense.

45

Lecture 5: Data Mining—Exploration and Prediction

&ODVVL¿FDWLRQDQG3UHGLFWLRQ z Perhaps the most common kind of task for data mining is using the data about an individual to predict some unknown characteristic about them. An “individual” here could be a person, or a river, or a bottle of wine, or anything else. The thing we’re trying to predict might be a categorical variable, such as what brand of car the person owns, or whether he or she would accept a particular offer to UH¿QDQFHWKHLUKRPHRUZKHWKHUKHRUVKHFRPPLWWHGIUDXGRQWKHLU federal income tax. When that’s the case, the prediction is often FDOOHGFODVVL¿FDWLRQ²VRUWLQJLQGLYLGXDOVLQWRWKHFODVVHVLQZKLFK we think they belong. z

On the other hand, we may be trying to predict a continuous variable, such as a person’s life expectancy, or the probability that D¿UPZLOOIDLOZLWKLQWKHQH[W¿YH\HDUV6RPHSHRSOHXVHWKHWHUP “prediction” only for such continuous variable cases, while some use prediction for sorting individuals into classes as well.

z

&ODVVL¿FDWLRQ WUHHV are a great data mining technique that we can picture as repeatedly subdividing rectangles into smaller and smaller rectangles until all of the points in any given rectangle fall into the same category of the output. If you’re working with continuous variables rather than classes, they’re usually called regression trees, but the idea is the same. This technique can be used on many different kinds of problems, it’s easy to explain to nontechnical people, the results are directly usable by the nonexpert, and it doesn’t care whether the data is normalized.

Important Terms FODVVL¿FDWLRQ: Using the information available about an individual to predict to which of a number of categories that individual belongs. FODVVL¿FDWLRQ WUHH $ GDWD PLQLQJ FODVVL¿FDWLRQ WHFKQLTXH 7KH DOJRULWKP repeatedly splits a subset of the individuals into two groups in a way that reduces total “impurity.”

46

heat map: A data visualization technique in which colors or shades are used to convey information about a variable in a chart, graph, or map. heteroscedastic: A collection of random variables is heteroscedastic if their standard deviations are not all equal. homoscedastic: A collection of random variables is homoscedastic if they all have the same standard deviation. Linear regression generally assumes that the error terms in the forecast are homoscedastic. normalizing: Also called standardizing. Linearly rescaling a variable to make its mean 0 and its standard deviation 1. This is done by taking each of the variable’s values, subtracting the mean of the variable, and then dividing the result by the variable’s standard deviation. prediction: In data mining, using the information available about an individual to estimate the value that it takes on some continuous output YDULDEOH6RPHSHRSOHXVHWKHWHUPWRLQFOXGHFODVVL¿FDWLRQDVZHOO SULQFLSDO FRPSRQHQWV DQDO\VLV 3&$ : A technique for reducing the number of variables in a data set by identifying a collection of linear combinations of those variables that capture most of the variation of a larger set of the original variables. These new variables can then be used in place of the larger set. regression tree$FODVVL¿FDWLRQWUHHZLWKDFRQWLQXRXVRXWSXW standard deviation: A measure of dispersion, it is the square root of the variance. variance: A commonly used statistical measure of the dispersion, or spread, of data. For a population, the variance is computed by deviation of each observation from the population mean, squaring those differences, and averaging the squares. For a sample, the same calculation is performed, but the result is multiplied by nní ZKHUHn is the sample size.

47

Suggested Reading Berry and Linoff, Data Mining Techniques. Dunham, Data Mining. Shmeuli, Patel, and Bruce, Data Mining for Business Intelligence.

Questions and Comments

Lecture 5: Data Mining—Exploration and Prediction

1. 7KHVSDP¿OWHULQWKHYLGHROHFWXUHKDGLQSXWYDULDEOHVWRGHWHUPLQH the output variable. Most of these record either the proportion of certain words in the email (“money” = 0.01 would mean that “money” made up 1% of the words) or the proportion of certain characters in the email (such as ! = 0.01 meaning that exclamation points made up 1% of the characters in the email). Notably different is the variable TOTCAPS, the total number of capital letters in the email. When PCA is done with WKH YDULDEOHV QRW EHLQJ QRUPDOL]HG WKH ¿UVW SULQFLSDO FRPSRQHQW LV dominated by TOTCAPS and captures 92.7% of the variance. When the YDULDEOHVDUHQRUPDOL]HGWKH¿UVWFRPSRQHQWFDSWXUHVRQO\RIWKH variation and is not dominated by anything. Can you explain why this kind of result is to be expected? Answer: The value of the TOTCAPS variable, when unnormalized, shows much more spread than the percentage variables, which run only between 0 and 1. As a result, the vast majority of the variation is in the TOTCAPS variable. When variables are normalized, each variable is on equal footing, with a variance of 1. Now the variation in each variable contributes equally to total variation.

2. $VHFRQGXVHRIWKHFODVVL¿FDWLRQWUHHWHFKQLTXHWKDWZHGHYHORSHGLVWR help with the reduction-of-variables problem. The variables appearing HDUO\ LQ RXU FODVVL¿FDWLRQ WUHH ZLOO SUREDEO\ EH LPSRUWDQW LQ IXUWKHU analysis of the problem, while those that appear quite late might be safely ignored.

48

'DWD0LQLQJIRU$I¿QLW\DQG&OXVWHULQJ Lecture 6

W

ith the mathematics of algorithms and the power of computers, data mining confers upon us an ability that could seem almost PDJLFDO WKH DELOLW\ WR ¿QG SDWWHUQV LQ D ZLOGHUQHVV RI GDWD WKDW E\DOOULJKWVVKRXOGGHI\RXUFRQFHSWLRQ6SHFL¿FDOO\GDWDPLQLQJKDVUDLVHG the business of anticipating customer preferences to a high art. For example, how does Amazon decide what offers to make to you? In this lecture, you will learn about some important tasks of data mining, including association and clustering. Training Data z Data mining involves using mathematical algorithms to extend our native ideas of pattern recognition. We can then use computers to implement those algorithms on large, multidimensional sets of data, OHWWLQJXV¿QGUHODWLRQVKLSVDQGSDWWHUQVWKDWWKHGDWDFRQWDLQV z

:LWK FODVVL¿FDWLRQ DQG SUHGLFWLRQ ZH ¿UVW VKRZ WKH DOJRULWKP a large set of training data. This data contains both the input and output values for a large collection of individuals. It’s like showing the algorithm a bunch of questions, along with their answers. The algorithm then generates a way of predicting the output value of new individuals, based on their input values.

z

Any such activity can be called prediction, although if the output YDOXH LV D FDWHJRU\²D FODVV²LW¶V XVXDOO\ FDOOHG FODVVL¿FDWLRQ All such procedures are called supervised learning, which merely means that you have to supply both the input and output values for each individual in the training data.

z

When we get our data—and in data mining, we often have a lot of it—we don’t use all of it for training. Instead, we partition it UDQGRPO\ LQWR WZR RU WKUHH JURXSV 7KH ¿UVW DQG DOPRVW DOZD\V

49

/HFWXUH'DWD0LQLQJIRU$I¿QLW\DQG&OXVWHULQJ

the largest) is used for training. But the second set is called the validation set, and it’s not used until after the data mining algorithm has been applied to the training data to generate its rules.

50

z

The performance of the model on the validation data is a fair test of the model’s quality, because that data had nothing to do with the model’s creation. This gives us a way to benchmark the results of different data mining algorithms and to evaluate which one looks best—the one that does best on the validation data.

z

However, selecting the model this way still doesn’t tell you how good you can expect it to be. The model you pick is the one that did best on the validation set, but would it do as well on other, new data? You picked the winning model because it was the best on the validation set, but maybe it just got lucky with that particular set.

z

That’s where the third set comes in: the testing set. If we have used the validation data to pick a best model, we’ve more or less poisoned its usefulness for assessing how the model does with new GDWD6RWKH¿QDOVWHSLVWRHYDOXDWHKRZZHOOWKHVHOHFWHGPRGHO does on brand new data—the testing data. Given a decently sized testing set, we should be able to use our model’s performance on it as a fairly reliable measure of its overall performance.

z

$SSO\LQJ WKLV WR FODVVL¿FDWLRQ WUHHV ZH EHJLQ E\ IROORZLQJ WKH procedure that continues by adding one node after another to the FODVVL¿FDWLRQWUHH²LQRWKHUZRUGVE\VXEGLYLGLQJUHFWDQJOHVDJDLQ DQGDJDLQ²XQWLOHYHU\SRLQWLVFRUUHFWO\FODVVL¿HG(YHQWXDOO\ZH reach the point where the original algorithm is no longer training to the real relationships between inputs and outputs; instead, it’s WUDLQLQJWR¿WWKHQRLVHLQWKHtraining set.

z

One way to handle this issue is called oversampling, and it’s used in particular when one of the classes of interest is especially uncommon in the data. The training set might include so few examples of the rare class of interest that a straightforward

application of data mining techniques would fail. The “best” rule for the technique might be to classify everyone as belonging to the largest class and no one to the class of interest. z

Oversampling includes a number of observations in the rare class disproportionate to its frequency in the population. You can do this by either taking a random sample (with replacement, if necessary) of individuals in the rare category and adding them to the training GDWDRUE\GXSOLFDWLQJHDFKUDUHFODVVHQWU\LQWKHWUDLQLQJVHWD¿[HG number of times.

z

In prediction problems, the quantity to be predicted is a continuous YDULDEOHUDWKHUWKDQDFDWHJRU\OLNHLQFODVVL¿FDWLRQSUREOHPV:H use the same kind of techniques, but we use different measures of error. Most software that can handle the one kind of problem can handle the other as well.

Association Rules z There are other interesting pattern recognition problems besides FODVVL¿FDWLRQDQGSUHGLFWLRQ'DWDPLQLQJDOVRGHDOVZLWKSUREOHPV of DI¿QLW\DQDO\VLV, or “what goes with what.” The recommender systems at online stores are an example. They use association rules. The system recognizes that you were looking at A and B, and then points out to you that people who liked them tend to be interested in C and D, too. z

In retail, association rules are sometimes called market basket analysis. If you know that people who buy these two products also tend to buy that one, you can use that information—for product placement, for promotional sales, and so on.

z

The strategy behind a basic association rule generator is actually pretty simple. In terms of retail sales, you have, for each customer in your database, a list of all of the items that the customer bought from you. You’re going to be looking for rules of the following form: “If a customer buys all of these, then I guess that he or she

51

/HFWXUH'DWD0LQLQJIRU$I¿QLW\DQG&OXVWHULQJ

wants these, too.” The game is probabilistic in nature, of course. You won’t be right every time, but you want your rules to give you useful guesses as to what might interest your customer. z

What do we need for a useful rule? First, there’s no point in considering an item—or group of items—that appear very rarely, at least not unless the combination is very important. So, you want to consider the support of a collection of items, which is merely the number of your customers—or fraction of your customers—who bought everything in the collection. If you try doing this in every possible way for a store with more than a small number of items, you’re going to run into trouble. A common work-around is to use the Apriori method.

z

Association rules software will scan the data to create frequent sets of items and then generate association rules from them and compute the FRQ¿GHQFH and lift index for each. Those with high values are good candidates for rules that the company may want to QRWHDQGXVH7KHUHLVGH¿QLWHO\DSODFHIRUKXPDQEHLQJVLQWKLV process; there are some rules that can be either discarded or merged with others.

Clustering z $VVRFLDWLRQUXOHVDUHORRNLQJDWLQIRUPDWLRQLQDYHU\VSHFL¿FZD\ that one set of attributes gives you reason to anticipate another set of attributes as well. But other techniques for looking for “things that go together” are based on the idea of clustering the data—of creating categories and then sensibly sorting the individual data SRLQWV LQWR WKHP 6RPHWLPHV WKLV LV D ¿UVW VWHS WR IXUWKHU DQDO\VLV that will be based on these clusters. z

52

The following two questions, of course, arise: what the categories should be and how we decide into which category an individual should be sorted. And the key idea to answering both of these questions is the same: distance.

z

7KH PRVW VWUDLJKWIRUZDUG ZD\ WR GH¿QH WKH GLVWDQFH EHWZHHQ two individuals is Euclidean distance, the straight-line distance separating the points in an nGLPHQVLRQDOVFDWWHUSORW7KDW¶VGLI¿FXOW WRYLVXDOL]HLQPRUHWKDQWKUHHGLPHQVLRQVEXWLW¶VQRPRUHGLI¿FXOW to compute with a souped-up version of the Pythagorean theorem from geometry.

z

7R ¿QG WKH GLVWDQFH EHWZHHQ WZR SHRSOH JLYHQ WKDW ZH KDYH LQIRUPDWLRQDERXWVL[FKDUDFWHULVWLFV$WKURXJK)ZH¿QGRXWKRZ much they differ on each of the six traits, square those differences, add them up, and take the square root.

z

$OWKRXJK WKLV LV WKH PRVW REYLRXV ZD\ WR GH¿QH WKH GLVWDQFH between two observations, there are some problems with it. There DUHRWKHUZD\VWRGH¿QHGLVWDQFHVXFKDVWKHstatistical distance, which both standardizes and accounts for correlation.

z

Before we cluster data points together, we have to answer a question: What do we mean by the distance between two clusters? 7KHGH¿QLWLRQZHFKRRVHGHSHQGVRQZKDWZHNQRZDERXWWKHGDWD

z

Suppose that we’re trying to track pollution levels. If a river is FRQWDPLQDWHGZHZRXOGH[SHFWWR¿QGSROOXWLRQFORVHWRWKHULYHU You don’t have to be close to all points on the river—just to some point on the river. Your distance from the river, then, would be measured by the distance to the closest point on the river. That’s RQHGH¿QLWLRQIRUWKHGLVWDQFHEHWZHHQWZRFOXVWHUV,W¶VFDOOHGWKH minimum distance, or the single linkage distance. If you use it, you can get clusters that tend to be ropey.

z

$VHFRQGSRVVLELOLW\JRHVWRWKHRSSRVLWHH[WUHPH