Simple Data Cleaning Tools and Methodologies

advertisement
From Data to Uncertainty
Principles of Data Quality
Albatrosses,
Kaikoura, New Zealand
Arthur D. Chapman
Australian Biodiversity Information Services
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
The Data Equation
Oceans of
Data
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Praia de Forte, Brazil
The
Data
Equation
Rivers of
Information
Ocean Biodiversity Informatics, Hamburg
Doubtful Sound, New Zealand
29 Nov 2004
The Data Equation
Streams
of
Knowledg
e
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Wasatch, Utah, USA
The Data Equation
Drops of
Understand
ing
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
(Nix 1984)
Taking Data to Information
Species
Data
Species
Data
Crab Florianopolis,
BrazilRock Cormorants
Wandering
Argentina
Orca, San
Albatros, NZ
Francisco
Armeria maritima
Brown
Algae,
Argentina
Algae,
New
Argentina
zealand
Corals, Australia
Environmental
Data
Information Temp
Rain
Range
Rain
June
Jan
Decisions
Policy
Conservation
Management
GIS
Data
Models
Decision
Support
Information
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
ThedoNeed
fortoModelling
Why
we need
use models?
Cont. USA
Brazil
Australia
Population:
Collections:
Plants:
Reptiles:
Mammals:
Insects:
Population: 172 million
Collections: 50 million
Plants:
70,000
Reptiles:
470-650
Mammals:
394
Insects:
?1.7 million
Population: 20 million
Collections: 35 million
Plants:
20,000
Reptiles:
850-890
Mammals:
305
Insects:
?220,000
292 million
0.5-1 billion
18,000
350
428
?150,000
Oceans
Population : ~ 0
Collections: ?10 million
Plants: ??
Vertebrates: ??
Invertebrates: ??
From OBIS 2004
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
What do we mean by ‘Data Quality’?
An essential or distinguishing characteristic
necessary for [spatial] data to be fit for use.
SDTS 02/92
The general intent of describing the quality of a
particular dataset or record is to describe the
fitness of that dataset or record for a particular use
that one may have in mind for the data.
Chrisman, 1991
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data quality - fitness for use?
Fitness for use
– Does species ‘A’ occur in Tasmania?
– Does species ‘A’ occur in National Park ‘y’
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
The Biological Data Domains
Errors can occur in any one of these
Plant and animal specimen data held in museums provide a vast
information resource, providing not only present day information on the
locations of these entities, but also historic information going back
several hundred years
(Chapman and Busby 1994).
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Loss of data quality
Loss of data quality can occur at many stages:
• At the time of collection
• During digitisation
• During documentation
• During storage and archiving
• During analysis and manipulation
• At time of presentation
• And through the use to which they are put
Don’t underestimate the simple elegance of quality
improvement. Other than teamwork, training, and
discipline, it requires no special skills. Anyone who wants to
can be an effective contributor.
(Redman 2001).
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Principles of data quality
The Vision:
• It is important for organizations to have a vision with
respect to having good quality data.
• As well as a vision, an organization needs a policy to
implement that vision.
• And a strategy for implementation
Experience has shown that treating data as a long-term asset and
managing it within a coordinated framework produces considerable
savings and ongoing value.
(NLWRA 2003).
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
The data quality vision
A Vision may involve
• Not reinventing information management wheels
• Looking for efficiencies in data collection and quality control
procedures
• Sharing of data, information and tools
• Using existing standards or developing new, robust standards
• Fostering the development of networks and partnerships
• Presenting a sound business case for data collection and
management
• Reducing duplication in data collection and data quality
control
• Looking beyond immediate use and examining requirements
of users
• Ensuring that good documentation and metadata procedures.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Strategies
Short term
- Data that can be assembled and checked over a 6-12
month period
Intermediate
- Data that can be entered over about an 18-month
period with small investment of resources
- Data that can be checked using simple in-house
methods
Long Term
- Data that can be entered and/or checked over a longer
time frame, using collaborative arrangements
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Information management chain
From: Chapman 2004
Assign responsibility for the quality of data to those who create them. If this is not possible,
assign responsibility as close to data creation as possible
(Redman 2001)
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles -1
• Planning is essential
– develop a vision, a policy and strategy
– Total Data Quality Management Cycle
Data ownership and custodianship not only confers rights to manage and
control access to data, it confers responsibilities
1 for its management, quality
control and maintenance. Custodians also have a moral responsibility to
superintend the data for Ocean
use by
future generations (Chapman 2004) 29 Nov 2004
Biodiversity Informatics, Hamburg
Data Cleaning Principles - 2
• Organising Data improves efficiency
– The organizing of data prior to data checking, validation and
correction can improve efficiency and considerably reduce
the time and costs of data cleaning.
– For example, by sorting data on location, efficiency gains
can be achieved through checking all records pertaining
to the one location at the same time, rather than going
back and forth to key references.
– Similarly, by sorting records by collector and date, it is
possible to spot errors where a record may be at an
unlikely location for that collector on that day.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 3
• Prevention is better than cure
– It is far cheaper and more efficient to prevent an error from
happening, than to have to detect it and correct it later. It is
also important that when errors are detected, that feedback
mechanisms ensure that the error doesn’t occur again during
data entry, or that there is a much lower likelihood of it reoccurring.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Asplenium bulbiferum, New Zealand
Data Cleaning Principles - 4
• Responsibility belongs to everyone
(collector, custodian and user)
– The principle responsibility belongs to the data custodian
– The collector has responsibility to respond to the custodian’s
questions when the custodian finds errors or ambiguities that
may refer back to the original information supplied by the
collector. These may relate to ambiguities on the label, errors in
the date or location, etc.
– The user also has a key responsibility to feed back to custodians
information on any errors or omissions they may come across,
including errors in the documentation associated with the data.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 5
• Partnerships improve efficiency
– By developing partnerships, many data validation processes
won’t need to be duplicated, errors will more likely be
documented and corrected, and new errors won’t be
incorporated by inadvertent “correction” of suspect records
that are not in error.
– Partnerships with:
• Data collectors
• Other institutions with duplicate collections
• Like-minded institutions developing tools, standards and
software
• Key data brokers (e.g. OBIS, GBIF)
• Data users (good feedback mechanisms)
• Statisticians and data auditors
Yours is not the only organization that is
dealing with data quality.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 6
• Prioritisation reduces duplication
– Prioritisation helps reduce costs and improves efficiency. It is
often of value to concentrate on those records where lots of
data can be cleaned at the lowest cost.
– For example, those that can be examined using batch
processing or automated methods, before working on the
more difficult records.
– By concentrating on those data that are of most value to
users, there is also a greater likelihood of errors being
detected and corrected.
Tierra del Fuego, Argentina
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Prioritising data quality procedures
• Focus on most critical data first
• Concentrate on discrete units (taxonomic, geographic, etc.)
• Ignore data that are not used or for which data quality
cannot be guaranteed
• Consider data that are of broadest value, are of greatest
benefit to the majority of users and are of value to the most
diverse of uses
• Work on those areas whereby lots of data can be cleaned
at the lowest cost (e.g. through use of batch processing).
Not all data are created equal, so focus on the most
important, and if data cleaning is required, make sure
it never has to be repeated (Chapman 2004).
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles -7
• Set targets and performance measures
– Performance measures are a valuable addition to quality
control procedures,
– They help an organization manage their data cleaning
processes.
– Performance measures may include statistical checks on the
data (for example, 95% of all records are within 1,000
meters of their reported position), on the level of quality
control (for example – 65% of all records have been checked
by a qualified taxonomist within the previous 5 years; 90%
have been checked by a qualified taxonomist within the
previous 10 years).
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 8
• Minimise duplication and re-working of data
– Duplication is a major factor with data cleaning in most
organizations.
– Many organizations add the geocode at the same time as
they database the record. As records are seldom sorted
geographically, this means that the same locations will be
chased up a number of times.
– By carrying out the georeferencing as a special
operation, records from similar locations can then be
sorted and then the appropriate map-sheet only has to
be extracted once.
– Some institutions also use the database itself to help
reduce duplication by searching to see if the location
may already have been georeferenced .
Ocean Biodiversity Informatics, Hamburg
Nothofagus antarctica, Argentina
29 Nov 2004
Data Cleaning Principles - 9
• Feedback is a two-way street
– Users of the data will inevitably carry out error detection, and
it is important that they feedback the results to the
custodians.
– It is essential that data custodians encourage feedback from
users of their data, and take the feedback that they receive
seriously.
– Data custodians also need to feed back information on
errors to the collectors and data suppliers where relevant.
– In this way there is a much higher likelihood that the
incidence of future errors will be reduced and the overall
data quality improved.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 10
• Education and training improves techniques
– Poor training, especially at the data collection and data entry
stages of the Information Quality Chain, is the cause of a
large proportion of the errors in primary species data.
– Good training of data entry operators can reduce the error
associated with data entry considerably, reduce data entry
costs and improve overall data quality.
Brown Algae, Argentina
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 11
• Accountability, Transparency and Audit-ability are
important
– Haphazard and unplanned data cleaning exercises are very
inefficient and generally unproductive.
– Within data quality policies and strategies – clear lines of
accountability for data cleaning need to be established.
– To improve the “fitness for use” of the data and thus their quality,
data cleaning processes need to be transparent and well
documented with a good audit trail to reduce duplication and to
ensure that once corrected, errors never re-occur.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Data Cleaning Principles - 12
• Documentation is the key to good data quality
– Without good documentation, it is difficult for users to determine
the fitness for use of the data and difficult for custodians to know
what and by whom data quality checks have been carried out.
– Documentation is generally of two types.
– The first is tied to each record and records what data checks
have been done and what changes have been made and by
whom.
– The second is the metadata that records information at the
dataset level.
– Both are important, and without them, good data quality is
compromised.
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Recording Accuracy and Error
Additional Accuracy Fields
–Preferably in meters (Point-Radius)
Documenting Validation tests
– Who
– What
– How
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Methods for geocode validation
•
•
•
•
•
Internal Database Checks
External Database Checks
Outliers in Geographic Space - GIS
Outliers in Environmental Space - Models
Statistical outliers
Ocean Biodiversity Informatics, Hamburg
Butterfly, Florida, USA
29 Nov 2004
Internal/External Database Checks
• Logical inconsistencies within the database
• Checking one field against another
– Text location vs geocode or District/State
• Checking one database against another
– Gazetteers
– DEM
– Collectors Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Magellanic Penguin, Argentina
Error
Error is inescapable and it should be recognised as a
fundamental dimension of data.
Chrisman 1991
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Geographic outliers - GIS
Country, State, named district, etc.
Gazetteer of Brazilian localities
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
35
1400
30
1200
Rainfall (mm)
Temperature (C)
Diva-GIS - Outlier
25
20
15
www.diva-gis.org
1000
800
600
400
• Reverse200jack-knifing technique
• Threshold0 value t = 0.95(n) +0.2
10
5
0
t
a
n
n
t
m
n
c
m
t
m
x
w
m
t
s
p
a
n
Ocean
t
c
l
q
t
w
m
q
Biodiversity
t
w
e
t
q
Informatics,
t
d
r
y
q
Hamburg
r
w
e
t
m
r
d
r
y
m 29
Nov
r
c
v
a
r
2004
r
w
e
t
q
r
d
r
y
q
r
c
l
q
r
w
m
q
CRIA-Data Cleaning
http://splink.cria.org.br/dc/
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Principal Components Analysis - FloraMap
Image from FloraMap (Jones and Gladkov 2001) showing use of Principal
Components Analysis to identify an outlier in Rauvolfia littoralis specimen data.
A. Principal Components Analysis
B. Specimen record. C. Mapped specimen.
Ocean Biodiversity Informatics, Hamburg
D. Climate profile
29 Nov 2004
Cumulative Frequency Curves - DivaGiS
Results from Diva-GIS showing the use of the Cumulative Frequency curve from BIOCLIM to identify possible geocoding errors in Rauvolfia
littoralis. A1 and A2 show possible outliers in climate space, B1 and B2 the corresponding mapped records. The Blue lines represent the
97.5 percentile
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Environmental Outliers
• Cumulative Frequency Curves
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Errors in data
Although most data gathering disciplines treat
error as an embarrassing issue to be
expunged, the error inherent in (spatial) data
deserves closer attention and public
understanding.
Chrisman, 1991
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Errors in data - 2
In general, error must not be treated
as a potentially embarrassing
inconvenience, because error
provides a critical component in
judging fitness for use.
Chrisman, 1991
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Mizodendrum sp., Argentina
Challengers
FutureFuture
Challengers
• Improved data quality
• Improved documentation of data
• Improved access to distributed data
• Improved methods for modelling in aquatic
(including marine) environments
• Decision Support Systems
Enlightened Policy / Decision Makers!!!
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Thank You…
Questions?
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Acknowledgements
•
•
•
•
•
•
•
•
•
•
Brazilian Biota/FAPESP Virtual Biodiversity Institute Program
Reference Centre for Environmental Information, Brazil (CRIA)
Global Biodiversity Information Facility (GBIF)
UNESCO
Wesleyan University, Connecticut, USA
Peabody Museum, Yale University, USA
ETI, Holland
UN Food and Agriculture Organization (FAO)
Environmental Resources Information Network, Australia (ERIN)
Commission on Data for Science and Technology (CODATA)
Ocean Biodiversity Informatics, Hamburg
29 Nov 2004
Download