Вы находитесь на странице: 1из 60

© 2008 Oracle. All rights reserved.

January 23, 2008

<Insert Picture Here>

Oracle Data Mining for


Targeted and Personalized Letter Generation
with BI Publisher
Mark F. Hornick
Senior Manager, Data Mining Technologies, Oracle
Agenda

• Traditional vs. intelligent form letters


• Oracle BI Publisher and Oracle Data Mining
• Scenarios and approach
• Overall process
• Response modeling
• Offer personalization
• Architecture
• Summary and Resources

© 2008 Oracle. All rights reserved.


Oracle BI Publisher +
Oracle Data Mining = “Green” Solutions
• U.S. catalog companies sent out 17 billion catalogs
last year - 59 for every man, woman and child in the
nation
• Magazine production in the U.S. uses more than 2.2
million tons of paper per year and the number is
increasing annually
• If everyone in the US was able to reduce their 10.8
pieces of junk mail received each week, we
could save nearly 100 million trees each year

© 2008 Oracle. All rights reserved.


Agenda

• Traditional vs. intelligent form letters


• Oracle BI Publisher and Oracle Data Mining
• Scenarios and approach
• Architecture
• Summary and Resources

© 2008 Oracle. All rights reserved.


Form Letters traditionally…

• All customers get the same form letter

or

• Vary form letters based on coarse customer segments

• Response rates are typically low: 0-4%

© 2008 Oracle. All rights reserved.


Form Letters traditionally…

• All customers get the same form letter, or


• Vary form letters based on coarse customer segments
• Manual (query-based) customer selection criteria is used to target
customers with one or more offers
• Segments based on (out)dated snapshot of customer information

• Response rates are typically low: 0-4%


• Content emailed/sent is irrelevant to over 95% of recipients!

© 2008 Oracle. All rights reserved.


“Intelligent” Form Letters…

• Dynamically predict which customers are likely to


respond based on latest customer information

• Dynamically predict which offers are likely to appeal


to individual customers

Enabled using BI Publisher with Oracle Data Mining

© 2008 Oracle. All rights reserved.


“Intelligent” Form Letters…

• Dynamically predict which customers are likely to


respond based on latest customer information
• Reduce campaign costs
• Avoid customer fatigue
• Dynamically predict which offers are likely to appeal
to individual customers
• “personalized” form letters
• Avoid making irrelevant offers
• Increase chances of customer response

Enabled using BI Publisher with Oracle Data Mining SQL scoring

© 2008 Oracle. All rights reserved.


Agenda

• Traditional vs. intelligent form letters


• Oracle BI Publisher and Oracle Data Mining
• Scenarios and approach
• Architecture
• Summary and Resources

© 2008 Oracle. All rights reserved.


Oracle BI Publisher
• Enterprise reporting solution to author, manage, and deliver all
types of highly formatted documents eliminating the need for
costly point solutions
• Reduces high costs associated with development,
customization and maintenance of business documents
• Increases efficiency of reports management
• Leverages familiar desktop tools where users can create and
maintain their own report formats based on data extracts from
diverse sources
• Rich Formatted Reports
• Partner Reports • Contracts
• Financial Statements • Checks
• Government forms • Labels
• Marketing materials • XML / EFT / EDI
e.g., Form Letters • Multiple destinations
© 2008 Oracle. All rights reserved.
Classic Reports
Customization
• 1 data set : 10 layouts : 30 languages >> 300 reports
• Requires high skilled engineers for customization

 High Maintenance Cost


 High Customization Cost
 Difficult Upgrade

© 2008 Oracle. All rights reserved.


BI Publisher Customization

• 1 data set : 10 layout templates : 30 translation files


• Business consultants use familiar
desktop tools for layout customization

XSL
Report
Templates

XML EFT

XML Data
Report Output
© 2008 Oracle. All rights reserved.
From Data to Destination

© 2008 Oracle. All rights reserved.


Oracle Data Mining
Option to Oracle 11g Database

• Sift through massive amounts of data to find


hidden patterns and discover new insights

• Understand the past, predict the future


• Take “reporting” to a whole new level

• Data Mining can provide valuable results…


• Identify factors more associated with a target attribute
(Attribute Importance)
• Predict individual behavior (Classification)
• Predict or estimate a value (Regression)
• Find profiles of targeted people or items
(Decision Trees)
• Segment a population (Clustering)
• Determine important relationships with the population
(Associations)
• Find fraud or rare “events” (Anomaly Detection)
© 2008 Oracle. All rights reserved.
Business Intelligence & Analytics
Query
and Reporting OLAP Data Mining
Extraction of Summaries, Knowledge discovery
detailed and trends and of hidden patterns
roll up data forecasts

“Information” “Analysis” “Insight & Prediction”

Who purchased What is the Who will buy a mutual


mutual funds in average fund in the next 6
the last 3 years? income of months and why?
mutual fund
buyers, by
region, by year?

© 2008 Oracle. All rights reserved.


Oracle Data Mining 11g
Oracle in-Database Mining Engine
Oracle 11g
11g DB
• Data Mining Functions (Server) Data Warehousing

• PL/SQL & Java APIs ETL

• Develop & deploy predictive analytics applications OLAP Statistics

• Integrate with other Oracle products Data Mining

• Oracle Data Miner (GUI)


• Simplified, guided data mining using wizards
• Predictive Analytics
• “1-click data mining” from a spreadsheet
• Wide range of DM techniques
• Classification & regression
• Clustering
• Anomaly detection
• Attribute importance
• Feature extraction
• Association rules (Market Basket analysis)
• Structured & unstructured data (text mining)

© 2008 Oracle. All rights reserved.


Oracle Data Mining
Algorithm Summary 11g
Technique Algorithm Applicability
Classification Logistic Regression (GLM) Classical statistical technique
Decision Trees Popular / Rules / transparency
Naïve Bayes Embedded app
Support Vector Machine Wide / narrow data / text

Regression Multiple Regression (GLM) Classical statistical technique


Support Vector Machine Wide / narrow data / text

Anomaly Detection One Class SVM Lack counter examples


Attribute reduction
Attribute Importance Minimum Description
Identify useful data
Length (MDL) Reduce data noise
Market basket analysis
Association Rules Apriori Link analysis

Clustering Hierarchical K-Means Product grouping


Text mining
Hierarchical O-Cluster Gene and protein analysis
Text analysis
Feature Extraction NMF Feature reduction
© 2008 Oracle. All rights reserved.
In-Database Analytics
Advantages Oracle 11g DB
Data Warehousing

• Data remains in the database at all ETL

times…with appropriate access security OLAP Statistics

control mechanisms—fewer moving parts Data Mining

• Straightforward inclusion within interesting


and arbitrarily complex queries
• Real-world scalability, performance, and availability
• Real-time scoring 2.5 million records scored in 6 seconds
on a single CPU system
• Pipeline results without costly materialization

© 2008 Oracle. All rights reserved.


Agenda

• Traditional vs. intelligent form letters


• Intro to Oracle BI Publisher and Oracle Data Mining
• Scenarios and approach
• Architecture
• Summary and Resources

© 2008 Oracle. All rights reserved.


Scenarios

• Response modeling
Generate standard form letter only to customers predicted
most likely to respond using up-to-the-moment data

• Offer personalization
Personalize a form letter with customer-specific offer(s),
also with up-to-the-moment data

© 2008 Oracle. All rights reserved.


How do we do this?

Obtain
Obtainhistorical
historicaldata
data Define
DefineDMDMpredictive
predictive
Build/test
Build/testDMDMmodels
models
with
withprevious,
previous,known
known SQL
SQLquery
queryinin
to
topredict
predict
customer
customerresponse
response BI
BIPublisher
Publisherover
over
offer
offerinterest
interest
outcomes
outcomes customer
customerdata
data

Write
Writeletter
letterinin Use
UseBI
BIPublisher
Publisher Evaluate
Evaluate
MS
MSWord
Wordwith
withcontent
content Server
Server post-campaign
post-campaign
from
fromquery
query to
togenerate
generateletters
letters results
results

© 2008 Oracle. All rights reserved.


Getting the historical data
Use previous campaign data or collect new…

Customer Population
for overall campaign

? ? ? ?
? ? ? ?
? ? Y N
? ? ? ? N
? ? ? ? Y N
? ? N
? ? ?
? ? Starter
Sample
? ?
Take a sample Run the starter
campaign
& get responses

© 2008 Oracle. All rights reserved.


Combine responses with predictor
attributes and build the model
Database Table with
Predictor Attributes
e.g., age, income, interests Target
Attribute
response, no-response

Model
If … then Y
Y N If … then N
N
Y N
N
Starter Sample Results

Combine results Build and test the model


with predictor attributes e.g. Decision Tree
automatically discovers rules

© 2008 Oracle. All rights reserved.


Take predictor attributes of remaining
customers and apply the model

Remaining Customers
Model Scores
Database Table with predict response, no-response
? ? Predictor Attributes If … then Y with probability

? ? If … then N
? ? ?
? ?
?

Apply the model

© 2008 Oracle. All rights reserved.


Rank order customers by
probability to respond or
filter with probability cutoff

select cust_id, name, address,…


from customer_scores
where rownum < 1100
order by prediction desc,
probability desc

select cust_id, name, address, …


from customer_scores
where prediction = ‘respond’
and probability > .6

© 2008 Oracle. All rights reserved.


Evaluate Post-Campaign Results
Y Y
N Y
N N Y
Y Y
N

Count accuracy of predictions


Scores Actuals
predicted response, no-response responded, did not respond,
revenue, cost
Y N
Y 400 25
N 75 600
Accuracy = (400+600) / 1100 = 90.9%
ROI = e.g., total profit / total cost * 100%

© 2008 Oracle. All rights reserved.


Scenario 1
Response Modeling
• Industry: Higher Education
• Domain: Endowment Growth
• Objectives:
• Increase success rate of alumni donations
• Reduce costs by targeting only likely donors
• Reduce alumni fatigue

© 2008 Oracle. All rights reserved.


Dataset
ALUMNI_RESPONSE

Target

© 2008 Oracle. All rights reserved.


Dataset
Predictors + Target

© 2008 Oracle. All rights reserved.


Build a Decision Tree Model using
Oracle Data Miner

© 2008 Oracle. All rights reserved.


Examine Decision Tree Rules
Corresponding to Alumni Segments/Profiles

© 2008 Oracle. All rights reserved.


Examine Test Results
for the Decision Tree Model

© 2008 Oracle. All rights reserved.


What benefit do we get
from the model?

Contact these first…

© 2008 Oracle. All rights reserved.


Obtain model name from ODMr

© 2008 Oracle. All rights reserved.


Predict alumni likely to respond

select alum_id, prediction, probability


from (select alum_id,
prediction(ALUMNI_RESPON66882_DT using *) prediction,
Check the query prediction_probability (ALUMNI_RESPON66882_DT using *)
probability
from ALUMNI_RESPONSE_9)
and where prediction = 1 and probability > .60

view the scores…

© 2008 Oracle. All rights reserved.


BI Publisher Server
Create a database connection

© 2008 Oracle. All rights reserved.


BI Publisher

Place ODM SQL


query here.
Returns only those
alumni likely to respond
to campaign.

© 2008 Oracle. All rights reserved.


Generate XML from BI Publisher
And load via Word Pluggin to build form letter…

© 2008 Oracle. All rights reserved.


Form Letter
to customers
expected to
respond

Upload template to
BI Publisher Server

© 2008 Oracle. All rights reserved.


Letters generated
only for those alumni
likely to respond to
campaign.

© 2008 Oracle. All rights reserved.


One step further…expected donation

• Build regression model to predict $$ donated

• Rank alumni by expected donation


Expected donation = probability of response * predicted donated amount

Select those Target Prediction


alumni who Attribute donation amount
donated amount
responded
Model
y = F(x1, x2, x3, …)

Build and test Apply


the model the model
using SVM or GLM to remaining alumni

© 2008 Oracle. All rights reserved.


Expected Donation SQL
select alum_id, (response_prob * pred_donation) expected_donation
from
(select alum_id,
prediction(ALUMNI_RESPON66882_DT using *) pred_response,
prediction_probability (ALUMNI_RESPON66882_DT using *) response_prob,
prediction (ALUMNI_DONAT43872_SVM using *) pred_donation
from ALUMNI_RESPONSE_9)
where pred_response = 1
order by expected_donation desc

© 2008 Oracle. All rights reserved.


Scenario 2
Offer Personalization
• Industry: Retail
• Domain: Electronics Direct Marketing
• Objectives:
• Provide most likely relevant offers to prospects
• Reduce costs by targeting only likely buyers
• Reduce customer fatigue

© 2008 Oracle. All rights reserved.


Dataset
CUSTOMER_OFFER

Target

© 2008 Oracle. All rights reserved.


Supplemental Data

OFFER

Name of predicted offer

© 2008 Oracle. All rights reserved.


Product Offer – Decision Tree

Predict which
offer a customer
will accept

© 2008 Oracle. All rights reserved.


Top offer for customer 100015

Check the query and view the results…

© 2008 Oracle. All rights reserved.


BI Publisher Server
Specify “Data Model” with DM query

Place ODM SQL


query here.
Generates one offer
per customer letter.

© 2008 Oracle. All rights reserved.


Create Word RTF file with
BI Publisher Template Builder

© 2008 Oracle. All rights reserved.


After loading XML data format,
drag and drop data field
to the Word document.

© 2008 Oracle. All rights reserved.


BI Publisher Server
• Upload RTF template to server
• View generated letters

Note that the customer ID would


be replaced with the customer name by
joining with a customer table
containing the actual names and addresses.

This execution produced 1200+ letters personalized


for each customer

© 2008 Oracle. All rights reserved.


Top 3 offers per Customer
An offer letter can contain more than one offer.
Test the query and view the results…then try in BI Publisher…

© 2008 Oracle. All rights reserved.


MyCompany
Offers
Offersincluded
includedspecifically
specificallyfor
for June 8, 2006
Offer Management Administration “Mr.
“Mr.Jones”.
Jones”.These
Theseare
are
1234 State Street
Bloomington, IN 88888
predicted
predicted to be of mostinterest
to be of most interest
to this customer.
to this customer.
Dear Mr. Jones,

We’d like to make you an offer...

• Save 20% on the TOWNE MP3 Player

256MB Music MP3 Player Supports MP3 & WMA music files Transfer and store any type of file Customizable user settings (Equalizer,
Backlight, Play Mode) PC and MAC compatible Connects to any USB port, no drivers required Includes AAA battery, carrying strap,
headphones and USB cable

• Save 30% on the PORTPLAY DVD Player

Progressive-scan playback: When connected to a digital TV, this DVD player delivers smooth, enhanced progressive-scan (480p)
images. Multi-format playback: DVD, DVD-R/RW, CD, CD-R/RW, MP3 and JPEG. A/V versatility: Component and S-video outputs
deliver superior picture quality. Digital audio output: Experience crisp digital soundtracks through a digital coaxial cable. Convenience
features: Aspect ratio selection, zoom/multi-angle functions, on-screen menu and more.

• Save 50% on the purchase of 3 or more DVDs

Popular titles available and USaveDVD.com

We hope to hear from you.

Sincerely,
Offer Management Administration

© 2008 Oracle. All rights reserved.


Agenda

• Traditional vs. intelligent form letters


• Intro to Oracle BI Publisher and Oracle Data Mining
• Scenarios and approach
• Architecture
• Summary and Resources

© 2008 Oracle. All rights reserved.


Architecture
Oracle DB EE
Data Miner
GUI Data Mining

BI Publisher Enterprise
Server
BIP Catalog

Template Builder
MS Word

© 2008 Oracle. All rights reserved.


Summary

• Oracle Data Mining combined with Oracle BI


Publisher facilitates targeted and personalized
offer letters

• Scoring customers using ODM models and SQL can


be quickly and easily integrated with BI Publisher to
prioritize offers on up-to-the-moment data

© 2008 Oracle. All rights reserved.


Resources

• Oracle BI Publisher
• http://www.oracle.com/technology/products/xml-publisher
• http://wiki.oracle.com/page/Oracle+BI+Publisher

• Oracle Data Mining


• http://www.oracle.com/technology/products/bi/odm/index.html
• http://wiki.oracle.com/page/Oracle+Data+Mining

• Oracle BIWA
• http://www.oraclebiwa.org

• Mark Hornick (mark.hornick@oracle.com)


Data Mining Technologies, Oracle

© 2008 Oracle. All rights reserved.


QUESTIONS
ANSWERS

© 2008 Oracle. All rights reserved.


© 2008 Oracle. All rights reserved.

Вам также может понравиться