The Question of Quality

Most of this presentation is based on the work of Marcos Gonçales

as cited in the references

Goals for this class

• Consider quality in digital libraries– How do we define quality– How do we measure quality– How does quality control impact a user?

Understanding Quality in a DL

• Quality indicators: proposed descriptions of quantities or observable variables that may be related to quality– “measures” = stronger term. Requires validation– Gonçalves et al provide analysis of quality conditions and

recommend specific quantities to be used.• Dimensions of quality• Proposed indicators• Application to DL concerns

Getting the data

• Where does the data come from?– Logging– Surveys– Focus Groups

• Know what information is needed, then choose the method most likely to provide the data.– More about the sources of data after we see what

we need to know.

What are we looking for?

• Consider that we are concerned about the quality of the following characteristics of a DL:– Data objects– Metadata– Collection– Catalog– Repository– Services

• What characteristics do we want each of those to have?

Dimensions of Quality

Dimensions of Quality

• Digital Object– Accessibility– Pertinence– Preservability– Relevance– Similarity– Significance– Timeliness

• Metadata Specification– Accuracy– Completeness – Conformance

• Collection– Completeness

• Catalog– Completeness– Consistency

• Repository– Completeness– Consistency

• Services– Composability– Efficiency– Effectiveness– Extensibility– Reusability– Reliability

What information do we need - related to Digital Objects

• Accessibility– What collection?– # of structured streams– Rights management metadata– Communities to be served

• Pertinence– Context– Information content– Information need

Information need - Digital Objects, continued

• Preservability– Fidelity (lossiness)– Migration cost– Digital object complexity– Stream formats

• Relevance– Feature frequency– Inverse document frequency– Document size– Document structure– Query size– Collection size

Information need - Digital Objects, continued

• Similarity– All the same features as in relevance– Also: citation/link patterns

• Significance– Citation/link patterns

• Timeliness– Age– Time of latest citation– Collection freshness

Information need - Metadata Specification

• Accuracy– Accurate attributes– # attributes in the record

• Completeness– Missing attributes– Schema size

• Conformance– Conformant attributes– Schema size

Information - Collection and Catalog

• Completeness of the Collection– Collection size– Size of an “ideal” collection

• Completeness of the Catalog– # of digital objects with no metadata

• Item level metadata

– Size of the collection

• Catalog Consistency– # of metadata specifications per digital object

Information about the Repository

• Completeness– # of collections

• Consistency– # of collections – Catalog/collection match

• How well do the catalogs match the collections?• Are the catalogs for all the collections at the

same level of detail?

Service Information Need

• Composability (ability to be combined to form new services)– Extensibility– Reusability

• Efficiency– Response time

• Effectiveness– Precision/recall (of search)– Classification

Service Information, continued

• Extensibility– # extended services– # services in the DL– # lines of code per service manager

• Reusability– # reused services– # services in the DL– # lines of code per service manager

• Reliability– # service failures– # accesses

Making more concrete

• Each of the measures listed gives an idea of the information need

• Exactly what do we measure?• How do we combine numbers obtained

to get a usable result?• Following pages describe specific

measures and formulas for combining those.

Digital Object Accessibility

• Basic requirement– If a user cannot access the DO, there is little point

in having it in the DL– Identified measures:

• Collection, # structured streams, rights management metadata, communities

– Say it another way:• Is it present in a collection in the repository?• Is there a service that can retrieve and display the

content?• Is the rights management open enough for access by

this user?

Digital Object Accessibility - formally

Define dox = a specific digital object

Accessibility = Acc(dox, acy) =– 0, if there is no collection C in the DL repository R

such that dox C– Otherwise, acc = (∑z struct_streams(dox) rz(acy))/ |

struc_streams(dox)|– where rz(acy)) is a rights management rule defined as

• 1, if – Z has no access constraints, or – Z has access constraints and acy cmz,

» Where cmz, Soc(1) is a community that has the right to access z; and

• 0, otherwise

This does not deal with accessibilty related to accessing the streams

An illustration

• NDLTD is the Networked Digital Library of Theses and Dissertations– Some institutions requre that all theses and

dissertations be stored in this DL– Student chooses how visible to make the

document.• Parts of the document may be visible while other parts

are not• The document, or parts of it, may be visible to a

restricted community.

Accessiblity case

• etdx is a specific electronic thesis or dissertation of interest

• acc(etdx) is– 0 if it is not in the collection– Otherwise (∑z struct_streams(etdx) rz(acy))/ |

struc_streams(dox)|• Where rz(acy) = 1

– if etdx is marked “world wide access” or etdx is marked “local institution only” and acy C where C is defined as identifiable members of the local institution

• = 0 otherwise

With the numbers

• An example from VT• For authors name beginning with A (219

entries):– Unrestricted ETDs: 164– Restricted ETDs: 50– Mixed ETDs: 5

• Percent unrestricted: 0.5, 0.5, 0.167, 0.1875, 0.6)

• Overall measure of accessibility outside VT:– (164 *1 + 50 * 0 + .5 + .5 + .167 + .1875 + .6)/219– 0.76

Solidifying Pertinence

• How do we measure something like pertinence?

• Relation between the information content of a digital object and the need of the user

• Depends on the user’s situation -- background, current context, etc.

Pertinence• Inf(doi) represents the information content of

digital object i• IN(acj) is the Information Need of actor (user) acj

• Context (acj, k) the combined effects of social factors that determine the pertinence of doi to acj

at time k• Two communities of actors

– Users whose information needs we try to satisfy– External Judges who are responsible for judging the

relevance of a document in response to a query. – Non overlapping groups

Pertinence formula

• Pertinence (doi, acj, k): Inf(doi) X IN(acj) X Context(acj, k) defined as– 1 if Inf(doi) is judged by acj to be informative with

regard to IN(acj) in context Context(acj, k)

– 0 otherwise

• Rather complex way to say that the information is relevant if either the user or a qualified independent judge says it is

Preservability

• Property of a digital object that describes its state relative to changes in hardware and software, representation format standards– Ex new recording technologies

(replacement of VHS video tapes by DVDs)– New versions of software such as Word or

Acrobat– New image standards such as JPEG 2000

Digital preservation techniques• Migration

– Transform from one format to another• Ex. Open the document in one format and save in another or do an

automated transformation

• Emulation– Reproducing the effect of the environment originally used to

display the material• Keep an old version of the software, or have new software that can

read the old format

• Wrapping– Keep the original format, but add enough human-readable

metadata so that it can be decoded in the future• Note that the material is not directly usable

• Refreshing– Copy the stream of bits from one location to another

• Particularly suitable for guarding against the physical deterioration of the medium

Most commonly used

Preservability issues• Obsolescence

– How out of date is the digital object?• Many versions of the software?• Old storage media?

– Difficult to migrate• Appropriate tools? Expertise?

• Fidelity – How different is the migrated version from the original?– Distortion = loss of information

• Preservability of a digital object in a digital library is a function of the fidelity of the migration and the obsolescence of the object

• Preservability(doi, dl) = (fidelity of migrating (doi, formatx, formaty), obsolescence(doi, dl))– Two values to reflect the two dimensions of the concept: fidelity and

obsolescence Miniclip Internet Archive

Preservability factors

• Capital direct costs– Software

• Developing software to create new versions of the object or obtaining licenses for new versions of the original software

– Hardware• For processing the migration and for storing the results

• Indirect operating costs– Monitoring digital objects for migration needs– Maintaining up-to-date intellectual property rights– Storage– Staff training

Calculating Obsolescence

• obsolence(doi, dl) = cost of converting/migrating the digital object, doi, within the context of a specific digital library

Calculating fidelity

• fidelity is the inverse of distortion.

fidelity(doi, formatx, formaty) =

1/(distortion(mp(formatx, formaty)) + 1.0) • One common measure of distortion

– mean squared error (mse)• Let {xn} be a stream of doi and {yn} be the converted stream• mse({xn}, {yn}) = ∑N

n-1(xn - yn)2 / N

Use mse for distortion:

fidelity(doi, formatx, formaty) = 1/(mse({xn}, {yn}) = ∑N

n-1(xn - yn)2 / N + 1.0)

No distortion: must yield a fidelity of 1.0

A Preservation ScenarioFrom Gonçales, adopted from one of his sources• Librarian learns that special collection of 1,000 digital images,

stored in TIFF v5.0, is in danger of obsolescence because the latest version of the display software does not support that version.

• Librarian decides to migrate all images to JPEG 2000, now the de facto image preservation standard, recommended by the Research Libraries Group (RLG)

• Librarian does search for options, finds a tool costing $500, that converts TIFF 5.0 to JPEG 2000

• About 20 hours needed to order, install, learn, apply the software to all images. Hourly rate of $66.60 per library employee.

• To save space, choose to use a compression rate that produces average mse = 8 per image.

• Preservability of each image = preservability (image-TIFF5.0, dl) = (1/9, ($500 +$66.60 *20)/1000) = (0.11, $1.83)

Both numbers are costs and lower is betterFidelity loss

Obsolescence costDistortion +1

Hourly rate * hours

# images

Relevance

• Relevance(d0i,q) =

= 1 if d0i is judged by an external judge to be relevant to query

q

= 0 otherwise• Measure of the distance between the vector representing

the object and the vector representing the object• The “external judge” requirement makes the measure

objective and independent of local contextual issues. Relevance has a consistency, independent of the momentary information need.

• Pertinence is a measure of usefulness within a particular information need.

Significance

• Significance is an expression of the absolute usefulness of a given digital object, independent of particular user needs.

• Citation records of objects in digital libraries offer one measure of significance. (This disadvantages the most recently obtained objects, since they have had less time to be cited by others.)

Look at ACM DL and the citation counts, for example.

Life Cycle and Quality

• The quality indicators relate to the core components of a digital library – creation, use, finding, distribution.

• Creation– Authoring, modifying– Describing, Organizing, Indexing

• Use– Access, filtering

• Finding (seeking)– Searching, Browsing, recommending

• Distribution– Storing– Archiving– Networking

Quality and Lifecycle - 2

Quality and Life Cycle - 3

• Note that some elements repeat– Timeliness is relevant to the content and to

the metadata that describes the content– Accessibility affects both usefulness and

distribution.

References

• Gonçalves, M. A., Moreira, B. L., Fox, E. A., and Watson, L. T. “Quality Model for Digital Libraries”.