Data Mining and Knowledge Discovery

1
Data Mining and Knowledge Discovery
By Matt Goliber and Jim Hougas
2
What is Data Mining?
Not like gold or diamond mining Mining of
knowledge from data Important to many
different fields A Part of Knowledge Discovery
in Databases (KDD)
3
The Process of Knowledge Discovery
Data cleaning and integration
Raw data
Data Warehouse
Data transformation, selection, and mining
Data transformation, selection, and mining
Pattern evaluation and knowledge presentation
Patterns
KNOWLEDGE!
4
Why is Data Mining useful?
We are data rich but information
poor -Internet -Intelligence Humans often
lack the ability to comprehend and manage the
immense amount of available and sometime
seemingly unrelated data
5
How long has this idea been around?
Late 60s and Early 70s Stanfords
Meta-DENDRAL (1970-76) -Extension of DENDRAL
Doug Lenat with AM (1976)
6
Meta-DENDRAL

• Extension of the DENDRAL (1965) program
• One of the first expert systems
• Interpreted mass spectra
• Meta-DENDRAL took the mass spectra of compound
  of known 3-D structure and formulated rules about
  the interpretation of the spectra
• Came up with known rules and some new ones!

7
Sample Mass Spec
ethyl 3-oxy-3-phenylpropanoate (ethyl
benzoylacetate)
8
AM

• Doug Lenat, 1976
• Name means nothing, stand alone
• AM was given sets, bags, ordered sets, and
  lists
• AM was also given operations to perform on
  these data sets
• Union, Intersection, ect
• Came up with ideas about counting, addition,
  multiplication, prime numbers, and Goldbachs
  conjecture
• AM thought that these were all uninteresting
• Liked maximally divisible numbers though

9
What next?
Not a whole lot Databases were not prevalent
enough, no great demand Did benefit from
machine learning research Beginning of the
1990s, The next area -Ranked as one of
the most promising research areas (NSF)
-Information explosion Early commercial
systems -Farm Journal -GM
10
Next Generation Techniques

• Decision Trees
• Each branch is a classification question
• Allows businesses to segment customers, products,
  and sales regions
• Questions organize the data
• Rule Induction
• All patterns are pulled from the data
• Accuracy and Significance are then added to them
• Help the user know how strong pattern is and
  likelihood of it occurring again
• Ex If bagels are purchased then cream cheese is
  purchased 90 of the time and this pattern occurs
  in 3 of all shopping baskets

11
Decision Trees vs. Rule Induction

• Decision Trees
• Many rules to cover same instance or
• no rule to cover an instance
• Rule Induction
• Always and only one rule
• Example
• Decision Trees use height and shoe size to
  determine size of person
• Rule Induction uses one or the other

12
Examples of Significant Developments

• Stock Market Advances (1991)
• Astrophysicists Doyne Farmer and Norman Packard
• Prediction company could predict stock market
  trends
• Bell Atlantic (1996)
• Consumer phone buying trends
• Rule Induction
• Advanced Scout (1997)
• Inderpal Bhandari assists NBA coaches
• Rule Induction
• Persuade 400,000 undecided voters (2004)
• MoveOn attemps to influence the election
• Decision Tree

13
Challenges

• Large Data Sets with High Complexity
• - One or the other is currently possible, but
  not both
• Expensive
• - Costs of Bell Atlantic (Experts are needed)
• - Cost for a two-day course in Las Vegas
  (1,300)
• - Software (100,000)

14
Research

• DARPA
• Defense Advance Research Projects Agency
• ACLU claims this is an invasion of privacy
• Decision Tree
• Uncovering Terrorists in public chat rooms
• Tracks the times that messages are sent
• Advanced Scout
• Bhandari is working on Advanced Scout for the NHL
• Rule Induction

15
Current State

• Out of the Lab
• Into Fortune 500 companies
• Automate Model Scoring
• Fingers are currently crossed in hopes that
  scoring by IT personnel is done correctly

16
Future States

• Utilizing Company Warehouses
• Data miners must take advantage of a million
  dollar warehouse that a company builds
• Effort Knob
• Low for quick model, high for quality model
• Computed Target Columns
• User could create a new target variable
• Ex finance information that a business has

17
Sources
http//web.media.mit.edu/haase/thesis/node54.html
SECTION00711000000000000000 http//smi-web.stanfo
rd.edu/projects/history.htmlMETADENDRAL http//ww
w.cs.cf.ac.uk/Dave/AI2/node151.html http//64.233.
161.104/search?qcacheQ6eMD9tEKwIJwww.cosc.brock
u.ca/Offerings/4P79/Week12.pptmeta-dendralhlen
http//laurel.actlab.utexas.edu/cynbe/muq/muf3_21
.html http//64.233.161.104/search?qcacheyft0cQ5
tZJQJwww.cs.uwaterloo.ca/shallit/Talks/cct.ps2
2fundamentaltheoremofarithmetic22computerdat
aminingprovehlen http//mathworld.wolfram.com/
GoldbachConjecture.html http//www.quantlet.com/md
stat/scripts/csa/html/node202.html http//www.thea
rling.com http//www.wired.com http//www.dmreview
.com http//www.ebscohost.com http//www.thearling
.com/text/dmtechniques/dmtechniques.htm http//www
.aaai.org/Library/Magazine/Vol13/13-03/vol13-03.ht
ml Data Mining Concepts and Techniques. Han J.
and Kamber M.