Characteristic Relational Patterns

Arne Koopman & Arno SiebesKDD 2009

Speaker: 王思元黎引得

Outline

2

Introduction Method Experiment Results Conclusion

Introduction

3

Relational database models Local models: frequent pattern mining Global models: probabilistic relational model

Charactering the database Combine patterns to form a global model

Relational Databases

4

KDD cup relational databases Genes Hepatitis Financial

ORDERorderID

loanID

accountID

Bank-ToAmountK-Symbol

LOANloanID

accountID

DateAmountDurationPaymentStatus

ACCOUNTloanID

accountID

FrequencyDate

CLIENTloanID

clientID

BirthyearSex

CARDcardID

loanID

dispositionID

TypeIssueDate

DISPOSITION

loanID

accountID

dispositionID

clientID

Type

1513 records

682 records

682 records 36 records

827 records

827 records

Local Model

5

Frequent pattern mining: too many patterns

ACCOUNTaccountID Date

10 200911 2009

LOANloanID accountID Duration100 10 12101 11 12102 11 24

ACCOUNTDate=2009

Patterns

ACCOUNTDate=2009

ACCOUNTDate=2009

LOANDuration=12

LOANDuration=12

LOANDuration=24

Global Model

6

Probabilistic Relational Models: local co-occurrence information lost

ACCOUNTaccountID Date

10 200911 2009

LOANloanID accountID Duration100 10 12101 11 12102 11 24

Probabilistic Relational Models

ACCOUNTDate=2009

LOANDuration=12

LOANDuration=24

0.67

0.33

Combined Model

7

Relational Code Table: compact and lossless description of the complete database

ACCOUNTaccountID Date

10 200911 2009

LOANloanID accountID Duration100 10 12101 11 12102 11 24

Code Table

ACCOUNTDate=2009

ACCOUNTDate=2009

LOANDuration=12

LOANDuration=12

LOANDuration=24

Outline

8

Introduction Method Experiments Conclusion

RDB-KRIMP RDB-KRIMP selects

patterns that describe the database well

Candidates are frequent relational patterns

Describing patterns are placed in a code table shortness code length

proportional to usagecode table

emptycodetable

database

database

select pattern

codetable

patterns

add to table

MDL

accept / reject

compress database

9

Outline

10

Introduction Method Experiments Conclusion

Experiments

11

Result for Different Minimum Support Value

12

100%

13

Comparison to Other Pattern Types

14

15

Outline

16

Introduction Method Experiments Conclusion

Conclusion Code tables describe the database while

preserving local information Code tables stay compact

Stay compact for low minimum support values Reductions up to 4 orders of magnitude

Richer patterns lead to better models Smaller encodings Better descriptions for database

17