Transcript No Slide Title - Computer Sciences User Pages

Data Quality Challenges in
Community Systems
AnHai Doan
University of Wisconsin-Madison
Joint work with Pedro DeRose, Warren Shen,
Xiaoyong Chai, Byron Gao, Fei Chen,
Yoonkyong Lee, Raghu Ramakrishnan, Jeff Naughton
Numerous Web Communities

Academic domains
– database researchers, bioinformatists
 Infotainments
– movie fans, mountain climbers, fantasy football
 Scientific data management
– biomagnetic databank, E. Coli community
 Business
– enterprise intranets, tech support groups, lawyers
 CIA / homeland security
– Intellipedia
Much Efforts to Build Community Portals


Initially taxonomy based (e.g., Yahoo style)
But now many structured data portals
– capture key entities and relationships of community
No general solution yet on how to build such portals
Cimple Project @ Wisconsin / Yahoo! Research
Develops such a general solution
using extraction + integration + mass collaboration
Maintain and add more sources
Jim Gray
Researcher
Homepages
*
**
*
Pages
* *
Group Pages
mailing list
Keyword search
SQL querying
Web pages
Conference
DBworld
Jim Gray
*
**
**
*
SIGMOD-04
**
*
Text documents
give-talk
SIGMOD-04
Question
answering
Browse
Mining
Alert/Monitor
News summary
DBLP
Mass collaboration
Prototype System: DBLife


Integrate data of the DB research community
1164 data sources
Crawled daily, 11000+ pages = 160+ MB / day
Data Extraction
Data Integration
Raghu Ramakrishnan
co-authors = A. Doan, Divesh Srivastava, ...
Resulting ER Graph
“Proactive Re-optimization
write
write
Shivnath Babu
coauthor
write
Pedro Bizarro
coauthor
advise
coauthor
Jennifer Widom
David DeWitt
PC-member
PC-Chair
SIGMOD 2005
advise
Provide Services

DBLife system
Mass Collaboration: Voting
Picture is removed if enough users vote “no”.
Mass Collaboration via Wiki
Summary: Community Systems

Data integration systems + extraction + Web 2.0
– manage both data and users in a synergistic fashion

In sync with current trends
– manage unstructured data (e.g., text, Web pages)
– get more structure (IE, Semantic Web)
– engage more people (Web 2.0)
– best-effort data integration, data spaces, pay-as-you-go
Numerous potential applications

But raises many difficult data quality challenges
Rest of the Talk

Data quality challenges in
1. Source selection
2. Extraction and integration
3. Detecting problems and providing feedback
4. Mass collaboration

Conclusions & ways forward
1. Source Selection
Maintain and add more sources
Jim Gray
Researcher
Homepages
**
*
*
Pages
* *
Group Pages
mailing list
Keyword search
SQL querying
Web pages
Conference
DBworld
Jim Gray
*
**
**
*
SIGMOD-04
**
*
Text documents
give-talk
SIGMOD-04
Question
answering
Browse
Mining
Alert/Monitor
News summary
DBLP
Mass collaboration
Current Solutions vs. Cimple

Current solutions
– find all relevant data sources
(e.g., using focused crawling, search engines)
– maximize coverage
– have lot of noisy sources

Cimple
– starts with a small set of high-quality “core” sources
– incrementally adds more sources
– only from “high-quality” places
– or as suggested by users (mass collaboration)
Start with a Small Set of “Core” Sources



Key observation: communities often follow 80-20 rules
– 20% of sources cover 80% of interesting activities
Initial portal over these 20% often is already quite useful
How to select these 20%
– select as many sources as possible
– evaluate and select most relevant ones
Evaluate the Relevancy of Sources

Use PageRank + virtual links across entities + TF/IDF
... Gerhard Weikum
G. Weikum
See [VLDB-07a]
Add More Sources over Time

Key observation: most important sources will
eventually be mentioned within the community
– so monitor certain “community channels” to find them
Message type: conf. ann.
Subject: Call for Participation: VLDB Workshop on Management of Uncertain Data
Call for Participation
Workshop on
"Management of Uncertain Data"
in conjunction with VLDB 2007
http://mud.cs.utwente.nl
...

Also allow users to suggest new sources
– e.g., the Silicon Valley Database Society
Summary: Source Selection

Sharp contrast to current work
– start with highly relevant sources
– expand carefully
– minimize “garbage in, garbage out”

Need a notion of source relevance
Need a way to compute this

2. Extraction and Integration
Maintain and add more sources
Jim Gray
Researcher
Homepages
**
*
*
Pages
* *
Group Pages
mailing list
Keyword search
SQL querying
Web pages
Conference
DBworld
Jim Gray
*
**
**
*
SIGMOD-04
**
*
Text documents
give-talk
SIGMOD-04
Question
answering
Browse
Mining
Alert/Monitor
News summary
DBLP
Mass collaboration
Extracting Entity Mentions


Key idea: reasonable plan, then patch
Reasonable plan:
– collect person names, e.g., David Smith
– generate variations, e.g., D. Smith, Dr. Smith, etc.
– find occurrences of these variations
ExtractMbyName
Union
s1 … sn
Works well, but can’t handle
certain difficult spots
Handling Difficult Spots

Example
– R. Miller, D. Smith, B. Jones
– if “David Miller” is in the dictionary
 will flag “Miller, D.” as a person name

Solution: patch such spots with stricter plans
ExtractMStrict
ExtractMbyName
Union
s1 … sn
FindPotentialNameLists
Matching Entity Mentions


Key idea: reasonable plan, then patch
Reasonable plan
– mention names are the same (modulo some variation)
 match
– e.g., David Smith and D. Smith
MatchMbyName
Extract Plan
Union
s1 … sn
Works well, but can’t handle
certain difficult spots
Handling Difficult Spots
MatchMStrict
DBLP: Chen Li
···
41. Chen Li, Bin Wang, Xiaochun Yang.
VGRAM. VLDB 2007.
···
38. Ping-Qi Pan, Jian-Feng Hu, Chen Li.
Feasible region contraction.
Applied Mathematics and Computation.
···
MatchMbyName
Extract Plan
Union
{s1 … sn} \ DBLP


Extract Plan
DBLP
Estimate the semantic ambiguity of data sources
– use social networking techniques [see ICDE-07a]
Apply stricter matchers to more ambiguous sources
Going Beyond Sources:
Difficult Data Spots Can Cover Any
Portion of Data
MatchMStrict2
MatchMStrict
Mentions that
Match “J. Han”
MatchMbyName
Extract Plan
Extract Plan
Union
{s1 … sn} \ DBLP
DBLP
Summary: Extraction and Integration

Most current solutions
– try to find a single good plan, applied to all of data

Cimple solution: reasonable plan, then patch
So the focus shifts to:
– how to find a reasonable plan?
– how to detect problematic data spots?
– how to patch those?



Need a notion of semantic ambiguity
Different from the notion of source relevance
3. Detecting Problems
and Providing Feedback
Maintain and add more sources
Jim Gray
Researcher
Homepages
**
*
*
Pages
* *
Group Pages
mailing list
Keyword search
SQL querying
Web pages
Conference
DBworld
Jim Gray
*
**
**
*
SIGMOD-04
**
*
Text documents
give-talk
SIGMOD-04
Question
answering
Browse
Mining
Alert/Monitor
News summary
DBLP
Mass collaboration
How to Detect Problems?

After extraction and matching, build services
– e.g., superhomepages
 Many such homepages contain minor problems
– e.g., X graduated in 19998
X chairs SIGMOD-05 and VLDB-05
X published 5 SIGMOD-03 papers
 Intuitively, something is semantically incorrect

To fix this, lets build a Semantic Debugger
– learns what is a normal profile for researcher, paper, etc.
– alerts the builder to potentially buggy superhomepages
– so feedback can be provided
What Types of Feedback?





Say that a certain data item Y is wrong
Provide correct value for Y, e.g., Y = SIGMOD-06
Add domain knowledge
– e.g., no researcher has ever published 5 SIGMOD
papers in a year
Add more data
– e.g., X was advised by Z
– e.g., here is the URL of another data source
Modify the underlying algorithm
– e.g., pull out all data involving X
match using names and co-authors, not just names
How to Make
Providing Feedback Very Easy?

“Providing feedback” for the masses
– in sync with current trends of empowering the masses

Extremely crucial in DBLife context
If feedback can be provided easily
– can get more feedback
– can leverage the mass of users


But this turned out to be very difficult
How to Make
Providing Feedback Very Easy?





Say that a certain data item Y is wrong
Provide correct value for Y, e.g., Y = SIGMOD-06
Add domain knowledge
Add more data
Provide form
interfaces
Modify the underlying algorithm
Provide a Wiki
interface
Critical in our
experience, but
unsolved
Unsolved, some
recent interest on
how to mass
customize software
See our IEEE Data Engineering Bulletin paper
on user-centric challenges, 2007
What Feedback
Would Make the Most Impact?
I have one hour spare time, would like to “teach” DBLife
– what problems should I work on?
– what feedback should I provide?
 Need a Feedback Advisor
– define a notion of system quality Q(s)
– define questions q1, ..., qn that DBLife can ask users
– for each qi, evaluate its expected improvement in Q(s)
– pick question with highest expected quality improvement
 Observations
– a precise notion of system quality is now crucial
– this notion should model the expected usage

Summary: Detection and Feedback

How to detect problems?
– Semantic Debugger
 What types of feedback &
how to easily provide them?
– critical, largely unsolved
 What feedback would make most impact?
– crucial in large-scale systems
– need a Feedback Advisor
– need a precise notion of system quality
4. Mass Collaboration
Maintenance and expansion
Jim Gray
Researcher
Homepages
**
*
*
Pages
* *
Group Pages
mailing list
Keyword search
SQL querying
Web pages
Conference
DBworld
Jim Gray
*
**
**
*
SIGMOD-04
**
*
Text documents
give-talk
SIGMOD-04
Question
answering
Browse
Mining
Alert/Monitor
News summary
DBLP
Mass collaboration
Mass Collaboration: Voting
Can be applied to numerous problems
Example: Matching
Dell laptop X200 with mouse ...
Mouse for Dell laptop 200 series ...
Dell X200; mouse at reduced price ...

Hard for machine, but easy for human
Challenges

How to detect and remove noisy users?
– evaluate them using questions with known answers

How to combine user feedback?
– # of yes votes vs. # of no votes
See [ICDE-05a, ICDE-08a]
Mass Collaboration: Wiki
Data
Sources
M
G
T

V1
W1
V2
W2
V3
W3
V3’
W3’
T3 ’
Community wikipedia
– built by machine + human
– backed up by a structured database
u1
Mass Collaboration: Wiki
Machine
<# person(id=1){name}=David J. DeWitt #>
Professor
<# person(id=1){title}=Professor #>
<strong>Interests:</strong>
<# person(id=1).interests(id=3)
.topic(id=4){name}=Parallel Database #>
Human
Human
<# person(id=1){name}=David J. DeWitt #>
<# person(id=1){title}=John P. Morgridge
Professor #>
<# person(id=1) {organization}=UW #>
since 1976
<strong>Interests:</strong>
<# person(id=1).interests(id=3)
.topic(id=4){name}=Parallel Database #>
David J. DeWitt
Interests:
Parallel Database
Machine
Machine
<# person(id=1){name}=David J. DeWitt #>
<# person(id=1){title}= John P. Morgridge
Professor #>
<# person(id=1){organization}=UW-Madison#>
since 1976
<strong>Interests:</strong>
<# person(id=1).interests(id=3)
.topic(id=4){name}=Parallel Database #>
<# person(id=1).interests(id=5)
.topic(id=6){name}=Privacy #>
David J. DeWitt
John P. Morgridge Professor
UW-Madison since 1976
Interests:
Parallel Database
Privacy
Sample Data Quality Challenges



How to detect noisy users?
– no clear solution yet
– for now, limit editing to trusted editors
– modify notion of system quality to account for this
How to combine feedback, handle inconsistent data?
– user vs. user
– user vs. machine
How to verify claimed ownership of data portions?
– e.g., this superhomepage is about me
– only I can edit it
See [ICDE-08b]
Summary: Mass Collaboration



What can users contribute?
How to evaluate user quality?
How to reconcile inconsistent data?
Additional Challenges







Dealing with evolving data (e.g., matching)
Iterative code development
Lifelong quality improvement
Querying over inconsistent data
Managing provenance and uncertainty
Generating explanations
Undo
Conclusions

Community systems:
– data integration + IE + Web 2.0
– potentially very useful in numerous domains

Such systems raise myriad data quality challenges
– subsume many current challenges
– suggest new ones

Can provide a unifying context for us to make progress
– building systems has been a key strength of our field
– we need a community effort, as always
See “cimple wisc” for more detail
Let us know if you want code/data