Master Data Management for Big Data

John R. Talburt, PhD, IQCP, CDMP

Black Oak Analytics, Inc, USA
MIT International Conference on Information Quality
Workshop June 21, 2016, Ciudad1 Real, Spain
© Black Oak Analytics
My Background
Currently
 Chief Scientist for Black Oak Analytics, Inc.
 Professor of Information Science and
Coordinator for the Information Quality Graduate Program
at the University of Arkansas at Little Rock (UALR)
Previously
 Business Leader for Data Research and Development at
Acxiom Corporation

© Black Oak Analytics 2

Talk Outline
Business Case for MDM
Technical Foundations of MDM
 Entity Resolution
 Entity Identity Information Management
 Master Data Management
The Need for Entity Resolution Analytics
Investing in Clerical Review for Continuous Improvement
Large-Scale MDM Using Distributed Processing

© Black Oak Analytics 3

 The Value Proposition for MDM

© Black Oak Analytics 4

The Business Case for MDM
Customer Satisfaction and Entity-Based Data Integration
Better Service
Reducing the Cost of Poor Data Quality
MDM as Part of Data Governance

© Black Oak Analytics 5

Customer Satisfaction
MDM has its roots in the customer relationship management (CRM)
industry.
The primary goal of CRM is to improve the customer’s experience and
increase customer satisfaction
The business motivation for CRM is to
 Increase customer retention rates
 Lower customer “churn rate”
 Gain new customers gained through social networking and referrals from
satisfied customers.
 Costs less to keep a customer than to acquire a new customer

© Black Oak Analytics 6

Better Service
Healthcare
 Improved clinical care, complete view patient encounters
 Improved medical research, find related cases
 The value proposition is “better quality of life”
Law Enforcement
 Many entity types- suspects, autos, airplanes, boats, phones, places, …
 Helps to bridge the many disparate and autonomous jurisdictions
 The value is more efficient and more effective investigation –
cases closed

© Black Oak Analytics 7

Reducing the Cost of Poor Data Quality
A major cause of data quality problems is “multiple source of the same
information produce different values for this information.”
 Lee, et al, “Journey to Data Quality”
A result of missing or ineffective MDM practices.
Taguchi’s Loss Function - the cost of poor data quality must be
considered not only in the effort to correct the immediate problem but
also include all of the costs from its downstream effects.
MDM is considered fundamental to an enterprise data quality program

© Black Oak Analytics 8

MDM as Part of Data Governance (DG)
DG is a program for managing information as an enterprise asset
DG provides a single-point of communication and control over
information in the enterprise
DG has created new management roles devoted to data and
information
 CDO, Chief Data Officer
 Data Stewards
 MDM and Reference Data Management (RDF) are regarded as essential
components of mature DG programs

© Black Oak Analytics 9

 Technical Foundations of MDM
Entity Resolution, Entity Identity Information Management, and MDM

© Black Oak Analytics 10

Three Related Concepts
Entity Resolution (ER)
Entity Identity Information Management (EIIM)
Master Data Management (MDM)

ER EIIM MDM

© Black Oak Analytics 11

Entity Resolution (ER)
The process of determining whether two references in an information
system are referring to the same real-world object or to different
objects (Talburt, 2011)
Record-linking
Record-deduplication
Data matching
Co-reference problem
Semantic resolution

If they refer to same real-world object, they are said to be “Equivalent”

© Black Oak Analytics 12

Which belong together?
J. Talburt
UALR T00020975 John Talburt
Wood Sorrell Pt, LR,AR
xxx-xx-1519
J Talburt
White Dr, Batesville, AR

John Talburt
700 Pierce, LR, AR
Jerry Talburt

John Talburt
Hazelton Rd, Pea Ridge, AR

John Talburt
J R Talburt DOB 3/1961
Old Country Ln, Cabot, AR

© Black Oak Analytics 13

Entity Identity Information Management (EIIM)
An extension of ER in two dimensions
 Knowledge management
• Creating, storing, and managing the information that represents the identity of an
entity
• Entity Identity Structure (EIS)
 Temporal
• Maintain persistent entity identifiers over time, i.e. process to process
Essential for
 Effective master data management (MDM)
 Entity-based data integration

© Black Oak Analytics 14

Master Data Management (MDM)
MDM is a collection of
 Policies, Procedures, Services, and Infrastructure
To support the
 Capture, integration, and shared use
Of
 Accurate, timely, consistent, and complete
Master data

David Loshin, Master Data Management

© Black Oak Analytics 15

Hierarchy of Support
Master Data Management (MDM)

Policies (Data Governance) Processes and Procedures

Entity Identity Information

Management (EIIM)

Entity Resolution (ER)

© Black Oak Analytics 16

Most Common MDM Mistakes Organizations Make
Fail to quantitatively and systematically measure and improve Entity
Identity Integrity achievement (Lack of QC and Continuous
Improvement)
Apply QA processes at the sourcing step, but not at the linking step
(Partial QA – Lack of Review Indicators)
Failure to address the life cycle of entity identity information
The EIIM information architecture is inadequate
The EIIM process is embedded in other ETL processes

© Black Oak Analytics 17

Measuring Entity Identity Integrity
Linking Accuracy = (TP+TN)/(TP+FP+TN+FN)
False Negative Rate = FN/(TP+FN)
False Positive Rate = FP/(TN+FP)
D R = set of References |R|=N
L E
D = All pairs in R, |D|= N*(N-1)/2
E = Equivalent Pairs
L-E LE E-L L = Pairs Linked by Process
FP TP FN

TN

© Black Oak Analytics 18

Measurement Techniques
Truth set development
 Small, but precise and time consuming
Benchmarking over the same dataset
 Large and fast, but less precise
Stratified sampling of clusters by attribute entropy
 In between, gives reliable accuracy statistics

© Black Oak Analytics 19

Quality Assurance at the Linking Step
Good MDM systems should produce “clerical review indicators”
Clerical review indicators are signals from the system that false positive
or false negative errors might have been made for certain linking
decisions
Clerical review indicators are implemented as “exception reports” that
should be reviewed by true domain experts who can decide if the error
was made or not
If errors were made, the experts should be able to override the system
and make corrections – “continuous improvement”

© Black Oak Analytics 20

 MDM Life Cycle Management
The CSRUD Model

© Black Oak Analytics 21

CSRUD Model
Capture of Entity Identity Information
Store and Share Entity Identity Information
Resolve and Retrieve Entity Identifiers
Update Entity Identity Information
Dispose (Retire) Entity Identity Information

© Black Oak Analytics 22

Capture Phase in an EIMS
Clerical
Link
Entity Review
Index
References Indicators

Application System

EIM Service

ER:
Staging Identity Identity KB
Capture
Build the initial population
Rules of identities (EIS) to be
managed
© Black Oak Analytics 23
Store & Share Phase
The Identity Knowledgebase is the primary repository of identity
information and provides a central point of management
The knowledgebase comprises the set EIS that represent each identity
under management
EIS vary from system to system and use different formats, e.g. XML
structures, relational database rows.

© Black Oak Analytics 24 24

Update Phase (Automated)
Clerical
Link Review
Entity
Index Indicators
References

Application System

EIM Service

Staging ER: Identity Updated

Update Identity KB

Current Rules
Identity KB
© Black Oak Analytics 25 25
Update Phase (Manual)
Clerical
Review Visualization Assertions
Indicators Tool

Application System

EIM Service

ER:
Current Identity Updated
Identity KB Update Identity KB

Rules

© Black Oak Analytics 26

Resolve and Retrieve (Batch)
Link Confidence
Entity Indicator
Index
References

Application System

EIM Service

ER: Identity
Staging
Resolution Identity KB

Uses IKB Information, but

Rules does not change it

© Black Oak Analytics 27

Resolve & Retrieve Phase (Interactive)
Application

Identity Entity
Information Identifier

Application System
API
EIM Service

ER: Identity Does not alter the IKB

Current
Resolution Identity KB

© Black Oak Analytics 28

Dispose (Retire) Phase
Eventually, some identities will no longer be relevant or active with
respect to the application
EIS can be moved from the IKB into an archive leaving only a
placeholder in the IKB.
Beware of schema change!
 When the definition of EIS change, it can create a problem in the retrieval
of archived information

© Black Oak Analytics 29

Pair- and Cluster-level Review Indicators
Pair-Level
 In Boolean (deterministic) systems – “Soft rules”
 In Scoring (probabilistic) systems – “Review threshold”
Cluster-Level
 Cluster Entropy
 Conflict Rules & Rationality Checks

© Black Oak Analytics 30

Example: Rationality Check at the Cluster Level
Source 1 Passes QA checks at
Source Level
EHR
Cluster 123

Name: John Doe; Annual Check-up:2015/01/15; … Does not make

sense at the
Name: John Doe; Deceased: 2010/04/08; … Cluster Level

Source 2
Passes QA checks at Unfortunately, many
Source Level organizations only perform QA
at the record level, and not at
the Cluster level

© Black Oak Analytics 31

 MDM in the World of Big Data
New IT Paradigms

© Black Oak Analytics 32

New IT Paradigm of Big Data
Move processes to data, not data to processes
Ingest data first, then analyze and determine model, not
design model first and force data to fit
Parse and structure data on output, not on input
De-Normalized key-value pair data stores, not normalized entity-
relation schemas
Implicit, middleware parallelism, not explicit coding

© Black Oak Analytics 33

Entity Resolution is a (Noisy) Graph Problem
21 11 3 22 15 10
A F D J A F D J

8
5
1 12 4
B E G K 4 7
B E G K
16 20 19
Match Key
H
9 Generator 1
I H Match Key
2
C 2 Generator 2
18 6 Match Key
C I Generator 3
Simple Undirected Graph 13

Match Key Graph

© Black Oak Analytics 34
Goal: Find the Connected Components
21 11 3 22 15 10
A F D J

12
8
5 Through a process
4
1
B E
4
G K
7 called the
16 20 19 “Transitive Closure”
Match Key
9 Generator 1 of the graph
H Match Key
2
2 Generator 2
18 6 Match Key
C I Generator 3
13

Match Key Graph

© Black Oak Analytics 35
Pre-Resolution Transitive Closure in Hadoop M/R
Generate Entity Resolution
Match Keys on Graph
Components

D1 Iterative D2
Refs Generate Entity Resolution
Transitive Merge
Match Keys on Graph
Closure
Components
Process

Entity Resolution
Generate on Graph
Match Keys Components

© Black Oak Analytics 36

Post-Resolution Transitive Closure
Hadoop M/R
Entity Resolution on
Refs Match Key 1 EIS

Hadoop M/R Hadoop M/R

Entity Resolution on Transitive
Refs Match Key 2 EIS Closure of Final
Entity Identifiers EIS

Hadoop M/R
Entity Resolution on
Refs Match Key 3 EIS

© Black Oak Analytics 37

Incremental Transitive Closure
Hadoop M/R
Sort by
Refs Entity
Match Key 1 EIS1
Resolution

Transitive Closure of Hadoop M/R

Entity Identifiers Entity
Refs from EIS 1 and EIS2
Resolution
Match Key 2

Transitive Closure of
Hadoop M/R
Entity Identifiers
Refs Entity
from EIS 2 and Final EIS
Resolution
Match Key 3

© Black Oak Analytics 38

Questions and Discussion

© Black Oak Analytics 39