business intelligence and data miningdatamining.rutgers.edu/teaching/fall2008/cis/lecture9.pdf ·...

23
1 Dr. Hui Xiong Business Intelligence and Data Mining Rutgers University Learning Objectives • Understand the need for business intelligence systems. • Know the characteristics of reporting systems. • Know the purpose and role of data warehouses and data marts. Understandfundamentaldatamining techniques. • Know the purpose, features, and functions of knowledge management systems. The Need for Business Intelligence Systems • According to a study done at the University of California at Berkeley, a total of 403 petabytes of new data were created. • 403 petabytes is roughly the amount of all printed material ever written. – The printed collection of the Library of Congress is .01 petabytes. – 400 petabytes equals 40,000 copies of the print collection of the Library of Congress. The Need for Business Intelligence Systems (Continued) • The generation of all these data has much to do with Moore’s Law. • The capacity of storage devices increases as thei ot de ea e theircosts decrease. • Today, storage capacity is nearly unlimited. • We are drowning in data and starving for information. Figure 91 How big is an Exabyte? Source: Used with permission of Peter Lyman and Hal R. Varian, University of California at Berkeley. Figure 92 HardDisk Storage Capacity Source: Used with permission of Peter Lyman and Hal R. Varian, University of California at Berkeley.

Upload: dangnhu

Post on 21-Aug-2018

214 views

Category:

Documents


0 download

TRANSCRIPT

1

Dr. Hui Xiong

Business Intelligence and Data Mining

gRutgers University

Learning Objectives

• Understand the need for business intelligence systems.

• Know the characteristics of reporting systems.

• Know the purpose and role of data warehouses and data marts.

U d d f d l d i i h i• Understand fundamental data‐mining techniques.

• Know the purpose, features, and functions of knowledge management systems. 

The Need for Business Intelligence Systems• According to a study done at the University of 

California at Berkeley, a total of 403 petabytes of new data were created.

• 403 petabytes is roughly the amount of all printed material ever written.– The printed collection of the Library of Congress is .01 petabytes.

– 400 petabytes equals 40,000 copies of the print collection of the Library of Congress.

The Need for Business Intelligence Systems (Continued)• The generation of all these data has much to do with Moore’s Law.

• The capacity of storage devices increases as thei o t de ea etheir costs decrease.

• Today, storage capacity is nearly unlimited.

• We are drowning in data and starving for information.

Figure 9‐1 How big is an Exabyte?

Source: Used with permission of Peter Lyman and Hal R. Varian, University of California at Berkeley.

Figure 9‐2 Hard‐Disk Storage Capacity

Source: Used with permission of Peter Lyman and Hal R. Varian, University of California at Berkeley.

2

Business Intelligence Tools

• Tools for searching business data in an attempt to find patterns is called business intelligence (BI) tools.

• Reporting tools are programs that read data f i t f th t d tfrom a variety of sources, process that data, produce formatted reports, and deliver those reports to the users who need them.

Business Intelligence Tools• The processing of data is simple:

– Data are sorted and grouped.– Simple totals and averages are calculated.

• Reporting tools are used primarily for assessment– They are used to address questions like:

•What has happened in the past?•What is the current situation?•How does the current situation compare to the past?

Business Intelligence Tools (Continued)• Data‐mining tools process data using statistical 

techniques, many of which are sophisticated and mathematically complex.

• Data mining involves searching for patterns and relationships among data.

• In most cases data mining tools are used to make• In most cases, data‐mining tools are used to make predictions.

• For example, we can use one form of analysis to compute the probability that a customer will default on a loan. 

• Another way to distinguish the differences of reporting tools and data‐mining tools is :– Reporting tools use simple operations like sorting, grouping, 

and summing.– Data‐mining tools use sophisticated techniques.

Business Intelligence Systems

• An information system is a collection of hardware, software, data, procedures, and people.

• The purpose of a business intelligence (BI) system is to provide the right information tosystem is to provide the right information, to the right user, at the right time.

• BI systems help users accomplish their goals and objectives by producing insights that lead to actions.

Business Intelligence Systems (Continued)• A reporting tool can generate a report that shows a 

customer has canceled an important order.

• A reporting system, however, alerts that customer’s salesperson with this unwanted news, and does so in time for the salesperson to try to alter the customer’s decisiondecision.

• A data‐mining tool can create an equation that computes the probability that a customer will default on a loan.

• A data‐mining system uses that equation to enable banking personnel to assess new loan applications.

Reporting Systems

• The purpose of a reporting system is to create meaningful information from disparate data sources and to deliver that information to the proper user on a timely basis.

• Reporting systems generate information from• Reporting systems generate information from data as a result of four operations:– Filtering data– Sorting data– Grouping data– Making simple calculations on the data

3

Figure 9‐3 Trade Data for NDX.X (NASDAQ 100) Figure 9‐4 Report Based on Trade Data in Figure 9‐3

Components of Reporting Systems• A reporting system maintains a database of reporting metadata.

• The metadata describes the reports, users, groups, roles, events, and other entities involved in the reporting activity.

• The reporting system uses the metadata to prepare and deliver reports to the proper users on a timely basis.

Figure 9‐5 Components of a Reporting System

Figure 9‐6 Summary of Report Characteristics Report Type• In terms of a report type, reports can be static or dynamic.

• Static reports are prepared once from the underlying data, and they do not change.– Example, a report of past year’s salesp , p p y

• Dynamic reports: the reporting system reads the most current data and generates the report using that fresh data.– Examples are: a report on sales today and a report on current stock prices 

4

Report Type (Continued)

• Query reports are prepared in response to data entered by users.

• Online analytical processing (OLAP) reports allow the user to dynamically change the report 

igrouping structures.

Report Media

• Reports are delivered via many different report media or channels.

• Some reports are printed on paper, and others are created in a format like PDF whereby they can be printed or viewed electronically.a e p i e o ie e e e o i a y

• Other reports are delivered to computer screens.

• Companies sometimes place reports on internal corporate Web sites for employees to access.

Report Media (Continued)• Another report medium is a digital dashboard, which is an electronic display customized for a particular user.– Vendors like Yahoo! and MSN provide common examples.p

– Users of these services can define content they want‐say, a local weather forecast, a list of stock prices, or a list of news sources.

– The vendor constructs the display customized for each user.

Report Media (Continued)• Other dashboards are particular to an organization.

– The organization might have a dashboard that shows up‐to‐the‐minute production and sales activities.

• Alerts are another form of report.– Users can declare that they wish to receive notifications of 

events say via email or on their cell phonesevents, say, via email or on their cell phones.

• Reports can be published via a Web service.– The Web service produces the report in response to requests 

from the service‐consuming application.

Figure 9‐7 Digital Dashboard Example Report Mode• The report mode can be either push report or pull report.

• Organizations send a push report to users according to a preset schedule.– Users receive the report without any activity p y yon their part.

• Users must request a pull report.– To obtain a pull report, a user goes to a Web portal or digital dashboard and clicks a link or button to cause the reporting system to produce and deliver the report.

5

Functions of Reporting Systems

• Three functions of reporting systems are:– Authoring– Management– Delivery

• Report authoring involves connecting to data sources, creating the reporting structure, and formatting the report.

Report Management• The purpose of report management is to define who 

receives what reports, when, and by what means.

• Most report‐management systems allow the report administrator to define user accounts and user groups and to assign particular users to particular groups.

• Reports that have been created using the report‐authoring system are assigned groups and users.

Report Management (Continued)• Assigning reports to groups saves the administrator work. – When a report is created, changed, or removed, the administrator need only change the report assignments to the group.

– All of the users in the group will inherit the changes.g p g

• Metadata also indicates what channel is to be used and whether the report is to be pushed or pulled.– If the report is to be pushed, the administrator declares whether the report is to be generated on a regular schedule or as an alert.

Report Delivery• The report‐delivery function of a reporting system 

pushes reports or allows them to be pulled according to report‐management metadata.

• Reports can be delivered via an email server, Web site, XML Web services, or by other program‐specific meansmeans. 

• The report‐delivery system uses the operating system and other program security components to ensure that only authorized users receive authorized reports.

Report Delivery (Continued)• The report‐delivery system also ensures that push reports are produced at appropriate times.

• For query reports, the report‐delivery system serves as an intermediary between the user andserves as an intermediary between the user and the report generator.– It receives user query data, such as item numbers in an inventory query, passes the query data to the report generator, receives the resulting report, and delivers the report to the user.

Online Analytical Processing• Online analytical processing (OLAP) provides the 

ability to sum, count, average, and perform other simple arithmetic operations on groups of data.

• The remarkable characteristics of OLAP reports is that they are dynamic.ey a e y a i

• The viewer of the report can change the report’s format, hence, the term online.

6

Online Analytical Processing• An OLAP report has measures and dimensions.

• A measure is the data item of interest.– It is the item that is to be summed or averaged or otherwise processed in the OLAP report.

• A dimension is a characteristic of a measureA dimension is a characteristic of a measure.– Purchase data, customer type, customer location, and sales region are all examples of dimension.

Online Analytical Processing (Continued)

• With an OLAP report, it is possible to drill down into the data.– This term means to further divide the data into more detail.

• Special‐purpose products called OLAP servers have been developed to perform OLAP analysis.

A O A f• An OLAP server reads data from an operational database, performs preliminary calculations, and stores the results of those operations in an OLAP database.

Figure 9‐13 OLAP Family and Store Location by Store Type Figure 9‐14 Role of OLAP Server and OLAP Database

Data Warehouses and Data Marts

• Basic reports and simple OLAP analyses can be made directly from operational data.

• For the most part, such reports display the current state of the business; and  if there are a few missing values or small inconsistencies with the data, no one is too concernedtoo concerned.

• Operational data are unsuited to more sophisticated analyses, particularly, data‐mining analyses that require high‐quality input for accurate and useful results.

Data Warehouses and Data Marts (Continued)

• Many organizations choose to extract operational data into facilities called data warehouses and data marts, both of which are facilities that prepare, store, and manage data specifically for data mining and other analyses.

• Programs read operational data and extract, clean, and g p , ,prepare that data for BI processing.

• The prepared data are stored in a data‐warehouse database using data‐warehouse DBMS, which can be different from the organization’s operational DBMS.

7

Data Warehouses and Data Marts• Data warehouses include data that are purchased from 

outside sources.

• Metadata concerning the data, its source, its format, its assumptions and constraints, and other facts about the data is kept in a data‐warehouse metadata database.p

• The data‐warehouse DBMS extracts and provides data to business intelligence tools such as data‐mining programs.

Figure 9‐15 Components of a Data Warehouse

Figure 9‐16 Consumer Data Available for Purchase from Data Vendors Problems with Operational Data (Continued)• Inconsistent data are particularly common for data that 

have been gathered over time.– When an area code changes, for example, the phone number 

for a given customer before the change will not match the customer’s number after the change.

• Some data inconsistencies occur from the nature of the business activitybusiness activity.

• Nonintegrated data can cause problems when data comes from different management information systems.

Figure 9‐17 Problems of Using Transaction Data for Analysis and Data Mining Data Warehouses Versus Data Marts• The data warehouse takes data from the data 

manufacturers (operational systems and purchased data), cleans and processes the data, and locates the data on the shelves, so to speak, of the data warehouse.

• A data mart is a data collection, smaller than the data warehouse, that addresses a particular component or functional area of the business.

8

Data Warehouse Versus Data Marts (Continued)

• The data warehouse is like the distributor in the supply chain and the data mart is like the retail store in the supply chain.

• Users in the data mart obtain data that pertain to a particular business function from the data warehouse.p

• It is expensive to create, staff, and operate data warehouses and data marts.

Figure 9‐18 Data Mart Examples

Data Mining and Business Intelligence

Dr Hui Xiong

Knowledge Discovery in Data

Dr. Hui XiongRutgers University

• Lots of data is being collected and warehoused – Web data, e‐commerce– purchases at department/

grocery stores

Why Mine Data? Commercial Viewpoint

– Bank/Credit Card transactions

• Computers have become cheaper and more powerful

• Competitive Pressure is Strong – Provide better, customized services for an edge (e.g. in 

Customer Relationship Management)

Why Mine Data? Scientific Viewpoint• Data collected and stored at 

enormous speeds (GB/hour)– remote sensors on a satellite

– telescopes scanning the skies

– microarrays generating gene expression datap

– scientific simulations generating terabytes of data

• Traditional techniques infeasible for raw data

• Data mining may help scientists – in classifying and segmenting data– in Hypothesis Formation

9

Mining Large Data Sets ‐Motivation• There is often information “hidden” in the data that is not readily evident

• Human analysts may take weeks to discover useful information

• Much of the data is never analyzed at all

3 500 000

4,000,000

0

500,000

1,000,000

1,500,000

2,000,000

2,500,000

3,000,000

3,500,000

1995 1996 1997 1998 1999

The Data Gap

Total new disk (TB) since 1995 Number of

analysts

From: R. Grossman, C. Kamath, V. Kumar, “Data Mining for Scientific and Engineering Applications”

Scale of DataOrganization Scale of Data

Walmart ~ 20 million transactions/dayGoogle ~ 8.2 billion Web pagesYahoo ~10 GB Web data/hrNASA satellites ~ 1.2 TB/dayNCBI GenBank ~ 22 million genetic sequencesFrance Telecom 29 2 TB

“The great strength of computers is that they can reliably manipulate vast amounts

of data very quickly. Their great weakness is that they don’t have a clue as to what any

France Telecom 29.2 TBUK Land Registry 18.3 TBAT&T Corp 26.2 TB

Why Do We Need Data Mining ?• Leverage organization’s data assets

– Only a small portion (typically ‐ 5%‐10%) of the collected data is ever analyzed

– Data that may never be analyzed continues to be collected, at a great expense, out of fear that g psomething which may prove important in the future is missing.

– Growth rates of data precludes traditional “manually intensive” approach

Why Do We Need Data Mining?

• As databases grow, the ability to support the decision support process using traditional query languages becomes infeasible– Many queries of interest are difficult to state in a query language (Query formulation problem)query language (Query formulation problem)

– “find all cases of fraud”– “find all individuals likely to buy a FORD expedition”

– “find all documents that are similar to this customers problem”

(Latitude, Longitude)1

What is Data Mining?

•Many Definitions– Non‐trivial extraction of implicit, previously unknown and 

potentially useful information from data– Exploration & analysis, by automatic or semi‐automatic 

means, of large quantities of data in order to discover meaningful patterns

What is (not) Data Mining?

What is Data Mining?

What is not Data Mining?

– Look up phone number in phone directory – Check the dictionary for the meaning of a word

– Certain names are more prevalent in certain US locations (O’Brien, O’Rurke, O’Reilly… in Boston area)– Group together similar documents returned by search engine according to their context (e.g. Amazon rainforest, Amazon.com,)

10

Data Mining: Confluence of Multiple Disciplines

?

20x20 ~ 2^400 ≈ 10^120 patterns

Data Mining Applications

• Market analysis• Risk analysis and management• Fraud detection and detection of unusual patterns (outliers)p ( )

• Text mining (news group, email, documents) and Web mining

• Stream data mining• DNA and bio‐data analysis

Fraud Detection & Mining Unusual Patterns

• Approaches: Clustering & model construction for frauds, outlier analysis

• Applications: Health care, retail, credit card service, …– Auto insurance: ring of collisions – Money laundering: suspicious monetary transactions 

Medical insurance– Medical insurance• Professional patients, ring of doctors, and ring of references• Unnecessary or correlated screening tests

– Telecommunications: phone‐call fraud• Phone call model: destination of the call, duration, time of day or week.  Analyze patterns that deviate from an expected norm

– Retail industry• Analysts estimate that 38% of retail shrink is due to dishonest employees

– Anti‐terrorism

Data Mining and Business Intelligence

Tid Refund Marital Status

Taxable Income Cheat

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No

4 Yes Married 120K No

Data

Data Mining Tasks …

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes

11 No Married 60K No

12 Yes Divorced 220K No

13 No Single 85K Yes

14 No Married 75K No

15 No Single 90K Yes 1 0

Milk

11

• Finding groups of objects such that the objects in a group will be similar (or related) to one another and different from (or unrelated to) the objects in other groups

Inter-clusterIntra-

Clustering

cluster distances

are maximize

d

acluster

distances are

minimized

• Understanding– Group related documents 

for browsing – Group genes and proteins 

that have similar functionality

– Group stocks with similar 

Discovered Clusters Industry Group

1 Applied-Matl-DOWN,Bay-Network-Down,3-COM-DOWN, Cabletron-Sys-DOWN,CISCO-DOWN,HP-DOWN,

DSC-Comm-DOWN,INTEL-DOWN,LSI-Logic-DOWN, Micron-Tech-DOWN,Texas-Inst-Down,Tellabs-Inc-Down,

Natl-Semiconduct-DOWN,Oracl-DOWN,SGI-DOWN, Sun-DOWN

Technology1-DOWN

2 Apple-Comp-DOWN,Autodesk-DOWN,DEC-DOWN, ADV-Micro-Device-DOWN,Andrew-Corp-DOWN,

Computer-Assoc-DOWN,Circuit-City-DOWN, Compaq-DOWN, EMC-Corp-DOWN, Gen-Inst-DOWN,

Motorola-DOWN,Microsoft-DOWN,Scientific-Atl-DOWN

Technology2-DOWN

3 Fannie-Mae-DOWN,Fed-Home-Loan-DOWN, MBNA-Corp-DOWN,Morgan-Stanley-DOWN

Financial-DOWN

4 Baker-Hughes-UP,Dresser-Inds-UP,Halliburton-HLD-UP, Louisiana-Land-UP,Phillips-Petro-UP,Unocal-UP,

Oil-UP

Applications of Cluster Analysis

pprice fluctuations

• Summarization– Reduce the size of large 

data sets

4 Schlumberger-UP Oil UP

Use of K‐means to partition Sea Surface Temperature (SST) and Net Primary Production (NPP) into clusters that reflect the Northern and Southern Hemispheres. 

Clustering: Application 1

• Market Segmentation:– Goal: subdivide a market into distinct subsets of customers 

where any subset may conceivably be selected as a market target to be reached with a distinct marketing mix.

– Approach:• Collect different attributes of customers based on their• Collect different attributes of customers based on their 

geographical and lifestyle related information.• Find clusters of similar customers.• Measure the clustering quality by observing buying 

patterns of customers in same cluster vs. those from different clusters. 

Clustering: Application 2

• Document Clustering:– Goal: To find groups of documents that are similar to each other based on the important terms appearing in them.

– Approach: To identify frequently occurring terms in each document. Form a similarity measure based on the frequencies of different terms. Use it to cluster.

What is not Cluster Analysis?• Simple segmentation

– Dividing students into different registration groups alphabetically, by last name

• Results of a query– Groupings are a result of an external specificationGroupings are a result of an external specification– Clustering is a grouping of objects based on the data

• Supervised classification– Have class label information

• Association Analysis– Local vs. global connections

Notion of a Cluster can be Ambiguous

How many clusters? Six Clusters

Four ClustersTwo Clusters

12

Types of Clusterings• A clustering is a set of clusters

• Important distinction between hierarchicaland partitional sets of clusters 

• Partitional Clustering– A division data objects into non‐overlapping subsets (clusters) such that each data object is in exactly one subset

• Hierarchical clustering– A set of nested clusters organized as a hierarchical tree 

Partitional Clustering

Original Points A Partitional Clustering

Hierarchical Clustering

p4 p1

p3

p2 p4p1 p2 p3

Traditional Hierarchical Clustering Traditional Dendrogram

p4p1

p3

p2

p4p1 p2 p3

Traditional Hierarchical Clustering

Non-traditional Hierarchical Clustering Non-traditional Dendrogram

Traditional Dendrogram

Other Distinctions Between Sets of Clusters• Exclusive versus non‐exclusive

– In non‐exclusive clusterings, points may belong to multiple clusters.

– Can represent multiple classes or ‘border’ points

• Fuzzy versus non‐fuzzy– In fuzzy clustering, a point belongs to every cluster with some 

weight between 0 and 1weight between 0 and 1– Weights must sum to 1– Probabilistic clustering has similar characteristics

• Partial versus complete– In some cases, we only want to cluster some of the data

• Heterogeneous versus homogeneous– Clusters of widely different sizes, shapes, and densities

Types of Clusters

• Well‐separated clusters

• Center‐based clusters

• Contiguous clusters

b d l• Density‐based clusters

• Property or Conceptual

• Described by an Objective Function

Types of Clusters: Well‐Separated

• Well‐Separated Clusters: – A cluster is a set of points such that any point in a cluster is closer (or more similar) to every other point in the cluster than to any point not in the cluster. 

3 well-separated clusters

13

Types of Clusters: Center‐Based

• Center‐based– A cluster is a set of objects such that an object in a cluster is closer (more similar) to the “center” of a cluster, than to the center of any other cluster  

– The center of a cluster is often a centroid, the average of all the points in the cluster, or a medoid,average of all the points in the cluster, or a medoid, the most “representative” point of a cluster 

4 center-based clusters

Types of Clusters: Contiguity‐Based

• Contiguous Cluster (Nearest neighbor or Transitive)– A cluster is a set of points such that a point in a cluster is closer (or more similar) to one or more other points in the cluster than to any point not in the cluster.

8 contiguous clusters

Types of Clusters: Density‐Based

• Density‐based– A cluster is a dense region of points, which is separated by low‐density regions, from other regions of high density. 

– Used when the clusters are irregular or intertwined, and when noise and outliers are presentand when noise and outliers are present. 

6 density-based clusters

Types of Clusters: Conceptual Clusters

• Shared Property or Conceptual Clusters– Finds clusters that share some common property or represent a particular concept.

2 Overlapping Circles

Characteristics of the Input Data Are Important• Type of proximity or density measure

– This is a derived measure, but central to clustering  • Sparseness

– Dictates type of similarity– Adds to efficiency

• Attribute type– Dictates type of similarity

• Type of Data– Dictates type of similarity– Other characteristics, e.g., autocorrelation

• Dimensionality• Noise and Outliers• Type of Distribution

Tid Refund Marital Status

Taxable Income Cheat

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No

4 Yes Married 120K No

Data

Data Mining Tasks …

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes

11 No Married 60K No

12 Yes Divorced 220K No

13 No Single 85K Yes

14 No Married 75K No

15 No Single 90K Yes 1 0

Milk

14

Association Rule Discovery: Definition

• Given a set of records each of which contain some number of items from a given collection– Produce dependency rules which will predict occurrence of 

an item based on occurrences of other items.

TID Items

1 Bread, Coke, Milk2 Beer, Bread3 Beer, Coke, Diaper, Milk4 Beer, Bread, Diaper, Milk5 Coke, Diaper, Milk

Rules Discovered:{Milk} --> {Coke}

{Diaper, Milk} --> {Beer}

Rules Discovered:{Milk} --> {Coke}

{Diaper, Milk} --> {Beer}

Association Analysis: Applications

• Market‐basket analysis– Rules are used for sales promotion, shelf management, and 

inventory management

• Telecommunication alarm diagnosis– Rules are used to find combination of alarms that occur 

together frequently in the same time period

• Medical Informatics– Rules are used to find combination of patient symptoms 

and complaints associated with certain diseases

Application Deployment Challenge

Tid Refund Marital Status

Taxable Income Cheat

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No

4 Yes Married 120K No

Data

Data Mining Tasks …

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes

11 No Married 60K No

12 Yes Divorced 220K No

13 No Single 85K Yes

14 No Married 75K No

15 No Single 90K Yes 1 0

Milk

• Find a model  for class attribute as a function of the values of other attributes

Tid Employed Level of Education

# years at present address

Credit Worthy

1 Yes Graduate 5 Yes

Model for predicting credit worthiness

Employed

No Yes

Predictive Modeling: Classification

1 Yes Graduate 5 Yes 2 Yes High School 2 No 3 No Undergrad 1 No 4 Yes High School 10 Yes … … … … …

10

No Education

Number ofyears

Graduate { High school, Undergrad }

Yes No

> 7 yrs < 7 yrs

Yes

Number ofyears

No

> 3 yr < 3 yr

Classification Example

Tid Employed Level of Education

# years at present address

Credit Worthy

1 Yes Graduate 5 Yes 2 Yes High School 2 No

Tid Employed Level of Education

# years at present address

Credit Worthy

1 Yes Undergrad 7 ? 2 No Graduate 3 ? 3 Yes High School 2 ? … … … … …

10

TestSet

Training Set Model

Learn Classifier

3 No Undergrad 1 No 4 Yes High School 10 Yes … … … … …

10

15

• Predicting tumor cells as benign or malignant

• Classifying credit card transactions as legitimate or fraudulent

• Classifying secondary structures of 

Examples of Classification Task

Classifying secondary structures of protein as alpha‐helix, beta‐sheet, or random coil

• Categorizing news stories as finance, weather, entertainment, sports, etc

• Identifying intruders in the cyberspace

Classification: Application 1• Fraud Detection

– Goal: Predict fraudulent cases in credit card transactions.– Approach:

• Use credit card transactions and the information on its account‐holder as attributes.– When does a customer buy, what does he buy, how often he pays on time, etc

• Label past transactions as fraud or fair transactions. This forms the class attribute.

• Learn a model for the class of the transactions.• Use this model to detect fraud by observing credit card 

transactions on an account.

Classification: Application 2• Churn prediction for telephone customers

– Goal: To predict whether a customer is likely to be lost to a competitor.

– Approach:• Use detailed record of transactions with each of the past 

and present customers, to find attributes.– How often the customer calls, where he calls, what 

time‐of‐the day he calls most, his financial status, marital status, etc.

• Label the customers as loyal or disloyal.• Find a model for loyalty.

From [Berry & Linoff] Data Mining Techniques, 1997

Classification: Application 3

• Sky Survey Cataloging– Goal: To predict class (star or galaxy) of sky objects, 

especially visually faint ones, based on the telescopic survey images (from Palomar Observatory).

– 3000 images with 23,040 x 23,040 pixels per image.– Approach:– Approach:

• Segment the image. • Measure image attributes (features) ‐ 40 of them per 

object.• Model the class based on these features.• Success Story: Could find 16 new high red‐shift quasars, 

some of the farthest objects that are difficult to find!

From [Fayyad, et.al.] Advances in Knowledge Discovery and Data Mining, 1996

Classifying Galaxies

Early

Intermediate

Class: • Stages of Formation

Attributes:• Image features, • Characteristics of light

waves received, etc.

Late

Data Size: • 72 million stars, 20 million galaxies• Object Catalog: 9 GB• Image Database: 150 GB

Classification Techniques

• Base Classifiers– Decision Tree based Methods– Rule‐based Methods– Nearest‐neighborN l N k– Neural Networks

– Naïve Bayes and Bayesian Belief Networks– Support Vector Machines

• Ensemble Classifiers– Boosting, Bagging, Random Forests

16

Example of a Decision Tree

ID Home Owner

Marital Status

Annual Income

Defaulted Borrower

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No

Home Owner

Yes No

Splitting Attributes

4 Yes Married 120K No

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes 10

MarSt

Income

YESNO

NO

NO

MarriedSingle, Divorced

< 80K > 80K

Training Data

Model: Decision

Tree

Another Example of Decision Tree

MarSt

Home Owner

Income

NO

NO

Yes No

MarriedSingle,

DivorcedID Home

OwnerMarital Status

Annual Income

Defaulted Borrower

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No Income

YESNO

NO< 80K > 80K

There could be more than one tree that fits the same data!

3 No Single 70K No

4 Yes Married 120K No

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes 10

Decision Tree Classification Task

Learn Model

Tid Attrib1 Attrib2 Attrib3 Class

1 Yes Large 125K No

2 No Medium 100K No

3 No Small 70K No

4 Yes Medium 120K No

5 No Large 95K Yes

6 No Medium 60K No

7 Yes Large 220K No

8 No Small 85K Yes

9 N M di 75K N

Apply Model

9 No Medium 75K No

10 No Small 90K Yes 10

Tid Attrib1 Attrib2 Attrib3 Class

11 No Small 55K ?

12 Yes Medium 80K ?

13 Yes Large 110K ?

14 No Small 95K ?

15 No Large 67K ? 10

Decision Tree

Apply Model to Test Data

Home Owner

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

DefaultedBorrower

No Married 80K ? 10

Test DataStart from the

root of tree.

MarSt

Income

YESNO

NO

NO

MarriedSingle, Divorced

< 80K > 80K

Apply Model to Test Data

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

Defaulted Borrower

No Married 80K ? 10

Test Data

Home Owner

MarSt

Income

YESNO

NO

NO

MarriedSingle, Divorced

< 80K > 80K

Apply Model to Test Data

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

DefaultedBorrower

No Married 80K ? 10

Home Owner

MarSt

Income

YESNO

NO

NO

MarriedSingle, Divorced

< 80K > 80K

17

Apply Model to Test Data

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

Defaulted Borrower

No Married 80K ? 10

Home Owner

MarSt

Income

YESNO

NO

NO

MarriedSingle, Divorced

< 80K > 80K

Apply Model to Test Data

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

Defaulted Borrower

No Married 80K ? 10

Home Owner

MarSt

Income

YESNO

NO

NO

Married Single, Divorced

< 80K > 80K

Apply Model to Test Data

MarStNO

Yes No

Home Owner

Marital Status

Annual Income

Defaulted Borrower

No Married 80K ? 10

Home Owner

MarSt

Income

YESNO

NO

NO

Married Single, Divorced

< 80K > 80K

Assign Defaulted to “No”

Decision Tree Classification Task

Learn Model

Tid Attrib1 Attrib2 Attrib3 Class

1 Yes Large 125K No

2 No Medium 100K No

3 No Small 70K No

4 Yes Medium 120K No

5 No Large 95K Yes

6 No Medium 60K No

7 Yes Large 220K No

8 No Small 85K Yes

Apply Model

Model9 No Medium 75K No

10 No Small 90K Yes 10

Tid Attrib1 Attrib2 Attrib3 Class

11 No Small 55K ?

12 Yes Medium 80K ?

13 Yes Large 110K ?

14 No Small 95K ?

15 No Large 67K ? 10

Decision Tree

Decision Tree Induction

• Many Algorithms:– Hunt’s Algorithm (one of the earliest)– CARTID3 C4 5– ID3, C4.5

– SLIQ,SPRINT

Tid Refund Marital Status

Taxable Income Cheat

1 Yes Single 125K No

2 No Married 100K No

3 No Single 70K No

4 Yes Married 120K No

Data

Data Mining Tasks …

5 No Divorced 95K Yes

6 No Married 60K No

7 Yes Divorced 220K No

8 No Single 85K Yes

9 No Married 75K No

10 No Single 90K Yes

11 No Married 60K No

12 Yes Divorced 220K No

13 No Single 85K Yes

14 No Married 75K No

15 No Single 90K Yes 1 0

Milk

18

Deviation/Anomaly Detection• Detect significant deviations from normal behavior

• Applications:– Credit Card FraudCredit Card Fraud Detection

– Network Intrusion Detection

Anomaly Detection

• Challenges– How many outliers are there in the data?– Method is unsupervised

• Validation can be quite challenging (just like for clustering)– Finding needle in a haystackFinding needle in a haystack

• Working assumption– There are considerably more “normal” observations than “abnormal” observations (outliers/anomalies) in the data

Anomaly Detection Schemes • General Steps

– Build a profile of the “normal” behavior•Profile can be patterns or summary statistics for the overall population

– Use the “normal” profile to detect anomalies•A li b ti h h t i ti•Anomalies are observations whose characteristicsdiffer significantly from the normal profile

• Types of anomaly detection schemes– Graphical & Statistical‐based– Distance‐based– Model‐based

Graphical Approaches• Boxplot (1‐D), Scatter plot (2‐D), Spin plot (3‐D)• Limitations

– Time consuming– Subjective

Statistical Approaches• Assume a parametric model describing the distribution of the data (e.g., normal distribution) 

• Apply a statistical test that depends on – Data distributionParameter of distribution (e g mean variance)– Parameter of distribution (e.g., mean, variance)

– Number of expected outliers (confidence limit)

Limitations of Statistical Approaches

• Most of the tests are for a single attribute• In many cases, data distribution may not be known

• For high dimensional data it may be• For high dimensional data, it may be difficult to estimate the true distribution

19

Distance‐based Approaches• Data is represented as a vector of features

• Three major approaches– Nearest‐neighbor based– Density based– Clustering based

Nearest‐Neighbor Based Approach• Approach:

– Compute the distance between every pair of data points

– There are various ways to define outliers:•Data oi t fo hi h the e a e fe e tha•Data points for which there are fewer than pneighboring points within a distance D

•The top n data points whose distance to the kth nearest neighbor is greatest

•The top n data points whose average distance to the k nearest neighbors is greatest 

Density‐based: LOF approach• For each point, compute the density of its local 

neighborhood• Compute local outlier factor (LOF) of a sample p as the 

average of the ratios of the density of sample p and the density of its nearest neighbors

• Outliers are points with largest LOF valuep g

p2× p1

×

In the NN approach, p2 is not considered as outlier, while LOF approach find both p1 and p2 as outliers

Clustering‐Based

• Basic idea:Cluster the data into groups of different densityChoose points in small l t did tcluster as candidate outliers

Compute the distance between candidate points and non‐candidate clusters. ‐ If candidate points are far from all other 

non‐candidate points, they are outliers

KDD Process• Develop an understanding of the application domain 

– Relevant prior knowledge, problem objectives, success criteria, current solution, inventory resources, constraints, terminology, cost and benefits

• Create target data set– Collect initial data, describe, focus on a subset of variables,Collect initial data, describe, focus on a subset of variables, 

verify data quality

• Data cleaning and preprocessing– Remove noise, outliers, missing fields, time sequence 

information, known trends, integrate data

• Data Reduction and projection– Feature subset selection, feature construction, discretizations, 

aggregations

KDD Process

• Selection of data mining task– Classification, segmentation, deviation detection, link analysis

• Select data mining approach D i i d l• Data mining to extract patterns or models

• Interpretation and evaluation of patterns/models

• Consolidating discovered knowledge

20

Knowledge DiscoveryChallenges of Data Mining

• Scalability• Dimensionality• Complex and Heterogeneous Data• Data Quality• Data Ownership and Distribution• Privacy Preservation• Streaming Data• Data from Multi‐Sources 

Similarities Between Data Miners and Doctors 

Data Characteristics

Data Mining Techniques Medical Devices

Commercial and Research ToolsWEKA: http://www.cs.waikato.ac.nz/ml/weka/

SAS: http://www.sas.com/

Clementine:Clementine:  http://www.spss.com/spssbi/clementine/

Intelligent Miner http://www‐3.ibm.com/software/data/iminer/

Insightful Miner http://www.insightful.com/products/product.asp?PID=26

Textbooks Knowledge Management• Knowledge management systems concern the sharing 

of knowledge that is already known to exist, either in libraries of documents, in the heads of employees, or in other known sources.

• Knowledge management (KM) is the process of i l f i ll l i l d h icreating value from intellectual capital and sharing 

that knowledge with employees, managers, suppliers, customers, and others who need that capital.

21

Knowledge Management (Continued)• Knowledge management is a process that is supported by the five components of an information system.– Its emphasis is on people, their knowledge, and effective means for sharing that knowledge with others.

• The benefits of KM concern the application of knowledge to enable employees and others to leverage organizational knowledge to work smarter.

• KM preserves organizational memory by capturing and storing the lessons learned and best practices of key employees.

Content Management Systems• Content management systems are information 

systems that track organizational documents, Web pages, graphics, and related materials.

• Such systems differ from operational document systems in that they do not directly support business 

ioperations.

• KM content management systems are concerned with the creation, management, and delivery of documents that exist for the purpose of imparting knowledge.

Content Management Systems (Continued)• Typical users of content management systems are 

companies that sell complicated products and want to share their knowledge of those products with employees and customers.

• The basic functions of content management systems are h f hthe same as for report management systems: author, manage, and deliver.

• The only requirement that content managers place on document authoring is that the document has been created in a standardized format.

Content Management Problems• Documents may refer to one another or multiple 

documents may refer to the same product or procedure.– When one of them changes, others must change as well.

– Some content management systems keep semantic linkages among documents so that content g gdependencies can be known and used to maintain document consistency.

• Document contents are perishable.– Documents become obsolete and need to be altered, removed, 

or replaced.

• Multinational companies have to ensure document language translations.

Figure 9‐23 Document Management at  Microsoft.com (as of December 2003)

Source: microsoft.com/backstage/inside.htm (accessed February 2004). © 2003 Microsoft Corporation. All rights reserved.

Figure 9‐24 Reporting Services: United States

Source: Used with permission of Tom Rizzo of Microsoft Corporation.

22

Figure 9‐25 Reporting Services: China

Source: Used with permission of Tom Rizzo of Microsoft Corporation.

Content Delivery

• Almost all users of content management systems pull the contents.

• Users cannot pull content if they do not know it exists.– The content must be arranged and indexed, and a facility for 

searching the content devised.searching the content devised.

• Documents that reside behind a corporate firewall, however, are not publicly accessible and will not be reachable by Google or other search engines.– Organizations must index their own proprietary documents 

and provide their own search capability for them.

KM Systems to Facilitate the Sharing of Human Knowledge

• Nothing is more frustrating for a manager to contemplate than the situation in which one employee struggles with a problem that another employee knows how to solve easily.

• KM systems are concerned with the sharing not only of content, but also with the sharing of knowledge among humans.– How can one person share her knowledge with another?– How can one person learn of another person’s great idea?

KM Systems to Facilitate the Sharing of Human Knowledge (Continued)

• Three forms of technology are used for knowledge‐ sharing among humans:– Portals, discussion groups, and email– Collaborations systems– Collaborations systems– Expert systems

Portals– Employees can share ideas by posting knowledge on a Web portal whereby managers and employees can pull the knowledge from the portal.

Figure 9‐26 Technology Support of Sharing Human Knowledge

KM Systems to Facilitate the Sharing of Human Knowledge (Continued)

Discussion Groups– Discussion groups allow employees or customers to post questions and queries seeking solutions to problems they have.Oracle IBM PeopleSoft and other vendors support– Oracle, IBM, PeopleSoft, and other vendors support product discussion groups where users can post questions and where employees, vendors, and other users can answer them.

– Later, the organization can edit and summarize the questions from such discussion groups into frequently asked questions (FAQs).

23

KM Systems to Facilitate the Sharing of Human Knowledge (Continued)Discussion groups (continued)

– Basic email can also be used for knowledge‐sharing, especially if email lists have been constructed with KM in mind.

– Two human factors inhibit knowledge‐sharing.•Employees can be reluctant to exhibit their ignorance.

•Competition exists between employees.

– A KM application may be ill‐suited to a competitive group.•The company may be able to restructure rewards and incentives to foster sharing of ideas among employees.

KM Systems to Facilitate the Sharing of Human Knowledge (Continued)Collaboration Systems

– Collaboration systems are information systems that enable people to work together more effectively.

– The Internet can be used as a broadcast medium for speeches, panel discussion, and other types of meetings.

– Web broadcasts, because they are digital, can be readily saved and replayed at the viewer’s convenienceand replayed at the viewer s convenience.

– Web broadcasts can also be made interactive by combining them with discussion group bulletin boards that are live during the broadcast.

– Video conferencing is another popular form of IT‐supported meetings.• Video‐conferencing equipment is expensive and normally is located in selected sites in the organization.

Figure 9‐27 Net Meeting Graphic KM Systems to Facilitate the Sharing of Human Knowledge (Continued)Expert Systems

– Expert systems are created by interviewing experts in a given business domain and codifying the rules stated by those experts.

– Many expert systems were created in the late 1980sMany expert systems were created in the late 1980s and 1990s, and some of them have been successful.

– Expert systems suffer from three major disadvantages.•They are difficult and expensive to develop.•They are difficult to maintain.•They were unable to live up to the high expectations set by their name.