Enterprise Knowledge Management. The Data Quality Approach

491 Pages·2001·6.484 MB·English

by  David Loshin (Auth.)

PREFACE
While data quality problems are widespread, it is rare for an event to
take place that provides a high-profile example of how questionable
information quality can have a worldwide business effect. The 2000 US
Presidential election and the subsequent confusion around the Florida
recount highlights the business need for high quality data. The winner
of this election is in a position to influence economies around the world.
The uncertainty associated with the lack of a clear winner had an
immediate effect the day after the election when stock prices plum
meted. Whether it is unintuitive data presentation, questions about the
way information is aggregated, or the method by which information
policy regulates the use of data, valuable information quality lessons
regarding at least six data quality issues that can be learned from the
election.
POOR DATA REPRESENTATION
A poor decision with respect to data presentation resulted in voter con
fusion. The use of the butterfly ballot in Palm Beach County, FL is an
example how the presentation of information did not correlate to users'
expectations, leading to a large number of voting errors. In fact, many
voters later claimed that they were dismayed to learn they may have
voted for Pat Buchanan, a right-wing Conservative candidate, instead
of Al Gore, the Democratic party candidate.
XIII
XIV PREFACE
DATA VALIDATION
With no built-in mechanism to vaHdate the data before it enters the sys
tem, the use of punch cards and the "butterfly ballot" leads to problems
with vote validation. When using a punch card ballot, (which, accord
ing to the LA Times was used by more than 37 percent of registered
nationwide voters in 1996), the voter selects a candidate by poking out
the chad—^the perforated section that should be ejected when the hole is
punctured. The cards are read by a tabulation machine, which counts a
vote when it reads the hole in the card.
The validation issue occurs when the chad is not completely
ejected. The automated tabulation of both "hanging chads" (chads that
are still partially attached) and "pregnant chads" (chads that are
bulging but not punched out) is questionable, and so it is not clear
whether all votes are counted. What constitutes a valid vote selection is
primarily based on whether the tabulation machine can read the card.
In the case of recounts, the cards are passed through the reader multiple
times. In that process some of the hanging chads are shaken free which
leads to different tallies after each recount.
In addition, if someone mistakenly punches out more than one
selection, the vote is automatically nullified. It is claimed that 19,000
ballots were disqualified because more than one vote for president had
been made on a single ballot. This is an example where a policy to pre-
qualify the ballot before it is sent to be counted could be instituted.
Since the rules for what constitutes a valid vote are well described, it
should be possible to have a machine evaluate the punch card to deter
mine whether it is valid or not, and notify the voter that the ballot
would be invalidated before it is cast.
INVALID ANALYTICAL MODELS
For many people on election eve, it is customary to sit in front of their
televisions and watch as their favorite newscasters predict the alloca
tion of electoral votes. These predictions are based on the results of exit
polls and election results provided by an organization called the Voter
News Service (VNS), which is jointly owned by a coalition of news
companies in order to cut the cost of data collection. Typically, the VNS
feeds both vote counts and winner predictions to all the news media
simultaneously, which is why all the different broadcasters seem to pre
dict the winners all around the same time.
PREFACE XV
In the case of the 2000 election, the networks were led to predict
the winner of Florida incorrectly, not just once, but twice. The first
error occured because predicting elections is based on statistical models
generated from past voting behavior that 1) were designed to catch vote
swings an order of magnitude greater than the actual (almost final) tal
lies and 2) did not take changes in demographics into account. This
meant that the prediction of Gore's winning Florida was retracted
about 2 hours after it was made.
CONFLICTING DATA SOURCES
By 2:00AM the VNS (and consequently, the reporting organizations)
switched their allocation of Florida's electoral votes from Gore to Bush,
and declared Bush to have enough electoral votes to win the election.
However, a second retraction occured when actual vote tallies were dis
puted. While the VNS report indicated that Bush led in Florida by
29,000 votes, information posted on the Florida Board of Elections web
site indicated that Bush's lead was closer to 500 votes, with the gap nar
rowing quickly. A computer glitch in Volusia County led to an overesti-
mation of Bush's total by more than 25,000 votes.
EXPECTATION OF ERROR
According to Title IX, Chapter 102 of Florida law, "if the returns for
any office reflect that a candidate was defeated or eliminated by one-
half of a percent or less of the votes cast for such office...the board
responsible for certifying the results of the vote...shall order a recount
of the votes..."
This section of the law contains data-accuracy implication that
there is an expected margin of error of one-half of one percent of the
votes. The automatic recount is a good example where the threshold for
potential error is recognized and where there is defined governance
associated with a data quality problem.
TIMELINESS
Timeliness is an important aspect of information quality. We have come
to expect that the results of a national election are decided by the time
XVI PREFACE
we wake up the next day. Even in a close election when the results are
inconclusive, there are timeliness constraints for the reporting and certi
fication of votes.
DATA QUALITY ON THE PERSONAL LEVEL
My personal connection to data quality began at a very early age,
although I did not realize the connection until recently. My parents,
through family coercion, decided to give me the first name "Howard,"
although they had wanted to give me the name "David." So David
became my middle name, but even though they named me Howard, I
have always been called — by my parents and, subsequently, everyone
else — David.
Hence the source of the problem. This fixed-field database-oriented
world is ill prepared to deal with a person who is called by his middle
name. Everywhere you go, you are asked for your first name and middle
initial. So, officially, I am Howard, but I always go by the name David.
At school, at camp, at college, at work, filling out applications, opening
bank accounts, filing tax returns, and so on, I fill out my name in its
official form.
People try to bend the rules: "Put down David as your first name
and H as your middle initial." "Just use David." "Scratch out first name
and change it to middle name." Unfortunately, these ideas are too radi
cal for the poor data entry people, so I either end up as David H. Loshin
or I am stuck with being called Howard.
This really happened: At a doctor's office recently, the receptionist
insisted that if my name were officially Howard D. Loshin on my in
surance card, they were only going to call me Howard. I have three dif
ferent credit cards, all with different names on them. Not only that —
my last name, Loshin, sounds the same as "lotion," and I find my last
name is consistently misspelled: Lotion, Loshen, Loshion, Loshian.
When it comes to how my mail is addressed, I never know what to
expect — except for one thing: I will get at least two of every direct
marketing sales letter.
Despite my inherited data quality connection, it was not until a few
years ago that I found a new motivation with data quality. I was asso
ciated with a securities processing group at a large financial services
company, who were analyzing their accounts database. What was inter
esting was that up to that point they had considered their accounts as
PREFACE XVII
just that: accounts. Over a period of time, however, some people there
became convinced of the benefits of looking at the people associated
with those accounts as customers, and a new project was born that
would turn the accounts database inside out. My role in that project was
to interpret the different information paradigms that appeared in the
accounts database name and address field. For it turned out that a single
customer might be associated with many different accounts, in many dif
ferent roles: as an owner, a trustee, an investment advisor, and so forth.
I learned two very interesting things about this project. The first
was that the knowledge that can be learned from combining multiple
databases was much greater than from the sum total of analyzing the
databases individually. The second was the realization that the prob
lems that I saw at this organization were not limited to this company —
in fact, these problems are endemic and not only within the financial
industry but in any industry that uses information to run its businesses.
The insight that brought full circle the world of data quality was
this: Every business process that uses data has some inherent assump
tions and expectations about the data. And these assumptions and
expectations can be expressed in a formal way, and this formality can
expose much more knowledge than simple database schema and Cobol
programs.
So I left that company and formed a new company. Knowledge
Integrity Incorporated, (www.knowledge-integrity.com) whose purpose
is to understand, expose, and correct data quality problems. Our goal is
to create a framework for evaluating the impacts that can be caused by
low data quality, to assess the state of data quality within an enterprise,
to collect the assumptions and expectations about the data that is used,
and recast those assumptions and expectations as a set of data quality
and business rules. In turn, these rules are incorporated as the central
core of a corporate knowledge management environment, to capture
corporate knowledge and manage it as content.
This book is the product of that goal. In it, we elaborate on our
philosophy and methods for evaluating data quality problems and how
we aim to solve them. I believe that the savvy manager understands the
importance of high-quality data as a means for increasing business
effectiveness and productivity, and this book puts these issues into the
proper context. I hope the reader finds this book helpful, and I am cer
tainly interested in hearing about others' experiences. Please feel free to
contact me at [email protected] and let me know how
your data quality projects are moving along!
XVIII PREFACE
I have been collecting what I call "Data Quality Horror Stories"
and placing them on our corporate Web site (www.knowledge-
integrity.com/horror.htm). If you have any interesting personal experi
ences, or if you see news stories that demonstrate how poor data quality
has serious (or comical) effects, please e-mail them to me at Loshin®
knowledge-integrity.com.
I would like to thank the people who have helped make this book
possible. First and foremost, my wife, Jill, and my children, Kira and
Jonah, are always there when I need them. Ken Morton acted as acqui
sitions editor and general enabler. Thomas Park, who took over the
project from Ken, was invaluable in helping me get this project com
pleted. I thank Thomas Redman, with whom I worked for a short
period of time and consulted on some of the concepts in this book.
Thank you to Mary O'Brien, who read through early drafts of the pro
posal and was a big supporter.
Thanks also go to Bob Shelly, whose experienced eye validated my
content, and to the rest of the Morgan Kaufmann staff involved in this
project. Sheri Dean and Julio Esperas also provided significant help in
the preparation of the book.
I also must thank Justin Kestelyn of Intelligent Enterprise maga
zine, who has vetted some of my ideas by publishing abridged versions
of a few chapters. Thanks also go to Dennis Shasha at New York Uni
versity, who gave me the opportunity to teach this material as a special
topic graduate class at the Courant Institute. I also thank my wife's par
ents, Marty and Phyllis Fingerhut, who are two of my biggest support
ers. Last, I want to thank and remember my mother, Betty Loshin, who
was a source of inspiration and who passed away earlier this year.
1
INTRODUCTION
Without even realizing it, everyone is affected by poor data quality.
Some are affected directly in annoying ways, such as receiving two or
three identical mailings from the same sales organization in the same
week. Some are affected in less direct ways, such as the 20-minute wait
on hold for a customer service department. Some are affected more
malevolently through deliberate fraud, such as identity theft. But when
ever poor data quality, inconsistencies, and errors bloat both companies
and government agencies and hamper their ability to provide the best
possible service, everyone suffers.
Data quality seems to be a hazy concept, but the lack of data qual
ity severely hampers the ability of organizations to effectively accumu
late and manage enterprise-wide knowledge. The goal of this book is to
demonstrate that data quality is not an esoteric notion but something
that can be quantified, measured, and improved, all with a strict focus
on return on investment. Our approach is that knowledge management
is a pillar that must stand securely on a pedestal of data quality, and by
the end of this book, the reader should be able to build that pedestal.
This book covers these areas.
• Data ownership paradigms
• The definition of data quality
• An economic framework for data quality, including steps in build
ing a return on investment model to justify the costs of a data
quality program
• The dimensions of data quality
• Using statistical process control as a tool for measurement
ENTERPRISE KNOWLEDGE MANAGEMENT
• Data domains and mappings between those domains
• Data quality rules and business rules
• Measurement and current state assessment
• Data quality requirements analysis
• Metadata and policy
• Rules-based processing
• Discovery of metadata and data quality and business rules
• Data cleansing
• Root cause analysis and supplier management
• Data enhancement
• Putting it all into practice
The end of the book summarizes the processes discussed and the steps
to building a data quality practice.
Before we dive into the technical components, however, it is worth
while to spend some time looking at some real-world examples for
motivation. In the next section, you will see some examples of "data
quality horror stories" — tales of adverse effects of poor data quality.
1.1 DATA QUALITY HORROR STORIES
1.1.1 Bank Deposit?
In November of 1998, it was reported by the Associated Press that a
New York man allegedly brought a dead deer into a bank in Stamford,
Connecticut, because he was upset with the bank's service. Police say the
70-year-old argued with a teller over a clerical mistake with his checking
account. Because he was apparently unhappy with the teller, he went
home, got the deer carcass and brought it back to the branch office.
1.1.2 CD Mail Fraud
Here is a news story taken from the Associated Press newswire. The
text is printed with permission.
Newark — For four years a Middlesex County man fooled
the computer fraud programs at two music-by-mail clubs,
using 1,630 aliases to buy music CDs at rates offered only to
first-time buyers.
INTRODUCTION
David Russo, 33, of Sayerville, NJ, admitted yesterday that
he received 22,260 CDs by making each address — even if it
listed the same post office box — different enough to evade
fraud-detection computer programs.
Among his methods: adding fictitious apartment numbers,
unneeded direction abbreviations and extra punctuation
marks. (Emphasis mine)
The scam is believed to be the largest of its kind in the
nation, said Assistant U.S. Attorney Scott S. Christie, who
prosecuted the case.
The introductory offers typically provided nine free CDs
with the purchase of one CD at the regular price, plus ship
ping and handling. Other CDs then had to be purchased later
to fulfill club requirements. Russo paid about $56,000 for
CDs, said Paul B. Brickfield, his lawyer, or an average of
$2.50 each. He then sold the CDs at flea markets for about
$10 each. Brickfield said. Russo pleaded guilty to a single
count of mail fraud. He faces about 12 to 18 months in
prison and a fine of up to $250,000.
1.1.3 Mars Orb/Cer
The Mars Climate Orbiter, a key part of NASA's program to explore the
planet Mars, vanished in September 1999 after rockets were fired to
bring it into orbit of the planet. It was later discovered by an investiga
tive board that NASA engineers failed to convert English measures of
rocket thrusts to newtons, a metric system measuring rocket force, and
that was the root cause of the loss of the spacecraft. The orbiter
smashed into the planet instead of reaching a safe orbit.
This discrepancy between the two measures, which was relatively
small, caused the orbiter to approach Mars at too low an altitude. The
result was the loss of a $125 million spacecraft and a significant setback
in NASA's ability to explore Mars.
1.1.4 Credit Card Woes
After having been a loyal credit card customer for a number of years, I
had mistakenly missed a payment when the bill was lost during the
ENTERPRISE KNOWLEDGE MANAGEMENT
move to our new house. I called the customer service department and
explained the omission, and they were happy to remove the service
charge, provided that I sent in my payment right away, which I did.
A few months later, I received a letter indicating that "immediate
action" was required. Evidently, I had a balance due of $0.00, and because
of that, the company had decided to revoke my charging privileges! Not
only that, I was being reported to credit agencies as being delinquent.
Needless to say, this was ridiculous, and after some intense conver
sations with a number of people in the customer service department,
they agreed to mark my account as being paid in full. They notified the
credit reporting agencies that I was not, and never had been, delinquent
on the account (see Figure 1.1).
1.1.5 Open or Closed Account?
Three months after canceling my cellular telephone service, I continue
to receive bills from my former service provider indicating that I was
being billed for $0.00 — "Do not remit."
1.1.6 Business Credit Card
A friend of mine is the president of a small home-based business. He
received an offer from a major charge card company for a corporate
charge card with no annual fee. He accepted, and a short time later, he
received his card in the mail. Not long after that, he began to receive the
same offer from the same company, but those offers were addressed dif
ferently. Evidently, his name had been misspelled on one of his maga
zine subscriptions, and that version had been submitted to the credit
card company as a different individual. Not only that, his wife started
to receive offers too.
Six months later, this man still gets four or five mail offers per week
in the mail from the same company, which evidently not only cannot fig
ure out who he is but also can't recognize that he is already a customer!
1.1.7 Direct Marketing
One would imagine that if any business might have the issue of data
quality on top of its list, it would be the direct marketing industry. Yet, I