Models and interaction mechanisms for exploratory interfaces
1. COMO CAMPUS
Models and interaction
mechanisms for
exploratory interfaces
Luigi Spagnolo
luigi.spagnolo@polimi.it
1 Information and Communication Quality
2. Index
2
¨ PREVIEW: Online experimentation!
¨ Part I: navigation, search and exploration
¤ Break
¨ Part II: Faceted search: the model(s) and the
interaction
¨ Visualization issues will be covered into an other
lecture
4. Intro
4
¨ This lecture starts in a quite unusual way :-)
¨ To let you introduced with exploratory
interfaces you’ll take part to a research
experiment
¨ But don’t worry!
¤ It’s
not dangerous for your health :-)
¤ The questionnaire you’re asked to fill is
anonymous and the answers will not be graded
6. The application | 2
6
¨ The last version of a prototype built for the Italian Ministry of
Culture
¨ A map of exploring venues of archaeological interest in Italy
¨ According to three properties (facets):
¤ Kind of venue: museum, archaeological site and superintendence (a
local branch of the Ministry of Culture devoted to archeological
heritage management).
¤ Location: the venue location, at level of macro-area (Northern Italy,
Central Italy, eyc.), Italian region and Italian province.
¤ Civilization or Period: The ancient civilizations (Romans, Greeks, etc.)
or periods (e.g. Middle Ages, Bronze age) the venues are relevant to.
7. The application | 3
7
¨ The tag cloud:
¤ Tag size à the number of results that are relevant with respect to the period or
civilization in question.
¤ Text color à how much the percentage of results that are relevant for the period/
civilization deviates from an uniform distribution.
n Shades of green show a stronger positive correlation between the other selected filters (e.g.
the location and/or the venue type) and the civilization/period in question. Red instead shows
a negative correlation (the civilization/period is less significant with respect to other criteria
selected).
¤ Background color à w.r.t. the whole set of venues are relevant for the period/
civilization, which percentage of them are included in the results?
n Green shows a positive correlation, while red instead shows a negative correlation.
¤ E.g., for venues in a specific region only (e.g. Lombardy), a green tagindicates that
the given civilization was particularly relevant for that region.
¤ The green background shows instead that the civilization is peculiar of that region,
and is less likely to be found elsewhere.
8. The application | 4
8
¨ The map:
¤ At three levels: Italian region, Italian province, extact
location(s)
¤ The color of the circle à the specific type of venue
¤ The size of the circle à the number of items of that
type in that area
9. The experiment
9
¨ Go to http://tinyurl.com/exp-icq
¤ (or http://www.ellesseweb.com/mining/)
¨ You will find a page with two links:
1. The application
2. An online questionnarie (on Rational Survey)
¤ Keep both open on the browser
¨ Work individually (1 hour max)
¨ Answer with your opinions, without looking at other websites, just
at the ArchaeoItaly application
¤ Remember: the survey is anonymous, and there are no “correct
answers”!
¤ For any doubts, ask me!
10. 10 Part 1 | Navigation, search and exploration
11. Let’s start with a scenario
11
¨ Work in pairs
¨ Imagine to work as journalists for the
Horse Illustrated magazine
¨ You have to write an essay about
horses in art (and in particular in
painting) among the centuries.
¨ Find interesting information on the
website of the Louvre Museum
¤ http://www.louvre.fr/llv/commun/
home.jsp?bmLocale=en
12. Problems with the Louvre
12
¨ Artworks are separated by department (internal
“bureaucratic” classification) and by
provenience.
¨ It is not possible to search them together
(regardless of their age and country of origin)
by subject.
¨ There is no introductory content on the subject
that can guide the student in her search.
13. Content-intensive websites
13
¨ Also know as:
¤ Information-intensive
¤ Often Infosuasive = informative + persuasive
¤ Like ancient rhetoric: inform and persuade
¨ Mainly intended for:
¤ Learning, understanding, discovering, comparing
information
¤ Leisure and entertainment
14. Contents
14
¨ Text, multimedia (audio, video, images)
¨ Hypermedia = multimedia + hyperlinks
¨ Information involves subjective judgment
¤ Depends on the author and on the user
¤ Objective: “10km far from Como”, “the painting
was made in 1886”
¤ Subjective: “Near Como”, “the painting is
impressionist”
15. User experiences requirements | 1
15
¤ From the users’ point of view:
n Usability: usage is effective, efficient and satisfactory
n Findability: users can locate what they are looking for
n “At a glance” understandabity: users understand the
website coverage and can make sense of information
n Enticing explorability: users are compelled to “stay
and play” and discover interesting connections
among topics
16. User experiences requirements | 2
16
¤ From the stakeholders’ point of view:
n Planned
serendipity: promoting most important
contents so that users can stumble in them
n E.g. “Readers that purchased this book also bought…”
n Communication strengh and branding: the website
conveys the intended “message” and “brand” of the
institution behind it
n E.g. “we have the lowest prices”, “we are very
authorithative”, etc.
17. Information architecture
17
¤ Purpose: conceptually
organizing information
¤ Providing access to contents
n Index navigation (a)
n Guided navigation (b)
¤ Providing the possibility of moving
from a content to related ones
n Contextual navigation (c): cross-
reference links, semantic relationships
18. “Traditional” structure
18
¤ Taxonomy: hierarchy
of categories and
subcategories
n Sections and group of
contents are the
branches of the tree
n Contents are the leaves
¤ Cross-reference links
between nodes
19. An example
19
¤ Artworks of the month
Sitemap:
¤ Paintings
Top 10 masterpieces
Art n
n By artist
gallery n By artistic movement
website n By subject
¤ Sculptures
n ...
n By material
¤ Photographs
n ...
20. Problems/1
20
¨ What if I want to browse all artworks (regardless
their type) by artist?
¤ Classifications are “nested” in a fixed order
¤ Designers should choose which classification should
prevail (e.g. by type)
¨ What if I want to find “impressionists paintings
portraing animals”?
¤ I cannot combine multiple “sibling”classifications (e.g.
by style and by subject)
21. Problems/2
21
¨ As long as the website is small a good
taxonomy can satisfy user requirements
¨ For large websites
¤ (hundreds or thousand of pages)
¤ Indexed/guided navigation doesn’t scale
¤ Users can’t easily find what they want
¤ Users can’t make sense of all such information
22. Solutions?
22
¨ What do users do when navigation doesn’t work?
¤ They use search!
¤ Search arranges contents dynamically and automatically (in
a way not predefined by designers)
¨ But keyword-based search is not optimal
¤ No hints for users that have no clear idea of what looking
for
¤ Users must know how the information is described (e.e.
the specific jargon used)
¤ Just for retrieval/focalized search
¨ We need a better paradigm: Exploratory search
23. Exploratory search
23
¨ The model “query à results” is (too
much) simple
¨ Search is often like berry picking!
(Bates 1989)
¤ Users explore a corpus of contents
¤ They refine the query (again and
again) according to what they learn
¤ They pick information here and there,
piece by piece
24. From search to exploration
24
¨ From finding to
understanding
(Marchionini)
¤ Acquire knowledge
about a domain, its
jargon, the properties of
information items in it.
¤ Useful to (better)
understand what to look
for
¤ …but also to analyze a
dataset
25. Goals of exploratory applications
25
¨ Object seeking
¤ Identify the best object(s) whose features match user
requirements (e.g. purchasing a photocamera with concerns
regarding price, resolution, etc.)
¨ Knowledge seeking
¤ Expand the knowledge about a given topic and related
information (e.g. Leonardo Da Vinci and Italian Renaissance)
¨ Wisdom seeking
¤ Discover interesting relationships among features in a
information space/dateset (e.g. analysis of sales in Esselunga
chain stores, according to store location, type of article, price,
etc.)
¨ These goals can possibly coexist in the same application
26. Retrieval vs. exploration models
26
¨ Retrieval model: query + results
¤ Query can can be either:
n Free form (e.g. keyword based search)
n Structured (parametric search, e.g. Scholar advanced search)
n Guided (select data from a predefined set of choices)
¨ Exploration model:
¤ Query + results + refinements/feedback
¤ Query supported by self-adaptive structures for:
n Further filter results to a subset of them
n Summarizing the features shared by results
27. 27 Part 2 | Faceted search: model(s) and interaction
(Amazon’s Diamond search was one of the first e-commerce applications of faceted search)
28. Faceted search
28
¨ A exploratory search/navigation pattern based on
progressive filtering of results
¨ The user selects a combination of metadata values belonging
to several facets
¨ Each facet correspond to a particular dimension that
describes the content objects made available for search, e.g.
for an artwork:
¤ Subject: people portrayed, flowers and plants, abstract...
¤ Medium: painting, sculpture, photography...
¤ Technique: oil, watercolors, digital art...
¤ Style: impressionism, expressionism, abstractism...
¤ Location: Prado, Louvre, Guggenheim
29. Let’s see a pair of examples
29
¨ Two examples:
¤ http://orange.sims.berkeley.edu/cgi-
bin/flamenco.cgi/famuseum/Flamenco
¤ http://www.artistrising.com
¨ Try the same search we’ve
seen before: find horses in
art
¨ More examples at:
http://www.flickr.com/photos/
morville/collections/
72157603789246885/
30. Non just a matter of finding…
30
E.g. you can learn
that horses in art
are often found in
paintings
portraing soldiers
or warriors and
leaders
31. How the interaction works
31
¨ When the user chooses a
filter, the application
selects:
¤ The results: items that have
been “tagged” with the filter
and the other metadata
previously chosen
¤ The remaining filters:
metadata that combined
with the previous choices
can produce results
¨ The users can continue
narrowing results until they
options are available
32. A (generalized) formal model | 1
( terms
)
32
¨ Taxonomy: a pair T ,
¤ A set of concepts or T = {t1 ,t2 ,…,tn }
¤ The subsumption relation connecting narrower
terms (hyponyms) to broader concepts (hypernyms)
laptop computer
location : 'Como' location : 'Lombardy' location : 'Italy'
¤ Terminal concepts: terms not further specialized
(the “leaves”)
33. A (generalized) formal model | 2
33
¨ For faceted taxonomies concepts are given
in terms of property-value pairs
(restrictions):
¤ E.g. subject: “horse”, location: “Como”
¨ A query is any of: q1 and q2
¤ A restriction q = property : value
q1 or q2
¤ A conjunction, disjunction or negation of
(sub)queries not q
¤ Actually there are limitations in the way concepts
can be combined in current facet browser
implementations
34. A (generalized) formal model | 3
34
¨ Item description: an information item o ∈O is
described as a conjunction of restrictions
d ( o ) = subject :"horse" and style :"Impressionism" and …
¨ Extension of a query: the set of items in a
context O that match the query ext ( q ) = {o ∈O | d ( o) q} O
ext ( q1 and q2 ) ⊆ ext ( q1 ) , ext ( q2 ) ( )
tc t p ⇒ ext ( tc ) ⊆ ext t p
ext ( q1 ) , ext ( q2 ) ⊆ ext ( q1 or q2 )
ext ( not q ) ≡ ext ( ALL) ext ( q )
35. A (generalized) formal model | 4
35
¨ The result of a query is:
¤ Itsextension in the given information space extO ( q )
¤ The set of features shared by these results: i.e. all
the concepts that can be derived from the
descriptions of objects in extO ( q )
36. Query transformations
36
¨ Operations allowing to navigate from a state to
another of the exploratio
¤ Appending new restrictions to the query in conjunction
(zoom-in: from a wider to a narrower set of results)
¤ Adding alternatives in disjunction to the existent ones (zoom-
out: from a narrower to a wider set)
¤ Removing existing constraints (zoom-out again)
¤ Negating/excluding values
¤ Replacing a filter with another (shift)
¨ Implemented by hyperlinks (for conjunctive filters / shift),
check boxes (for disjunctions), etc.
37. How values are (usually) combined
37
¨ Filters belonging to different facets are combined in
conjunction
¤ E.g. “technique:oil” AND “style:impressionism”
¤ Filters belonging to the same facet are:
¤ Combined in conjunction if the facet admits more values at
the same time for each object
n E.g. “subject:people” AND “subject:animals”
n (both people and animals in the same picture)
¤ Combined in disjunction if the facet adimits only one value
n E.g. “location:Milan” OR “location:Como”
n (an object which is Como or in Milan)
38. Type of facets
38
¨ Single-valued (functional properties) vs. multi-valued
¨ Flat vs. hierarchical organization of values
¤ E.g. hierarchical: nation/region/province
¨ Subjective/arbitrary (properly named facets) vs. objective
(attributes)
¤ A date, a location, a price are examples of objective data
¤ “Topic”, “Audience”, “Artistic movement”, “importance” are
examples of subjective information
¤ Assigning/using a value involves some kind of judgment and
interpretation and is influenced by cultural and personal
backgrounds
39. Type of facet values
39
¨ Terms (strings of text) ¨ Sortable and comparable?
¤ Taxonomies, controlled ¤ We can say that
vocabularies value1<=value2<=…<=valueN?
¤ User-defined tags ¤ E.g. Dates, magnitudes, scales of
(folksonomies) judgment, quantitative data
n e.g. “sufficient”<“excellent”,
¤ From data-mining
10€<100€, “Monday”<“Friday”
¨ Numerical values and dates ¤ Ranges [value1, value2]
¨ Boolean values (yes/no) n E.g. User is allowed to search for
events from 01/06 to 31/08
¤ E.g. “Available for buying?”,
“original?”, “still living?” ¤ Classes of values
n e.g. for price: 0-10€, 11-20€,
¨ Even shades of color, 21-50€, 51-100€, …
shapes, etc... n The way we define classes is arbitrary
and depend on domain
40. Benefits of faceted search
40
¨ Easy and natural almost like “traditional” browsing
¨ With respect to keyword-based search users have hints
¤ Users can more easily make sense of information (if supported by
good interfaces)
¤ …and learn about the context by interacting with it
¨ Users can freely combine multiple classifications according to their
wishes
¤ In traditional browsing, when you reach a terminal concept you can’t
refine further
¤ With faceted search, you can continue refining with related concepts
¨ Navigation is safe: frustrating “no results found” searches avoided
¤ Only concepts that have been used to classify the current set of
results are diplayed
41. Limitations
41
¨ It works well only with structured data
¨ Faceted search does not provide a ranking of
results
¤ For “object seeking” tasks it might be a limitation
¤ It may be better to compute the “distance” with
respect to an “optimal” solution à otimization task
¨ Other limitations are discussed in the following
slides on advanced issues
43. Full Boolean queries | 1
43
¨ How to achieve something like this?
“Given a budget of 250,000 euros,
I’m interested in a flat with at least 4
rooms and not central heating in the
centre, or an house with at least 5
rooms in the suburbs”
44. Full Boolean queries | 2
44
¨ Foci (Ferré et al.) the set of sub-expressions in the semantic
tree of the query
¨ ( )
A query is a pair q,φ , where q is an arbitrary combination of
filters and φ is one of its foci
¤ The focus is used to select the subquery at which the new filter
should be appended (or the transformation should be applied)
¤ …But also to “inspect” different points of view of information
¤ The main focus represents the “whole” query
45. Semantic faceted search
45
¨ We can filter items, but how can we filter facet values?
¤ E.g. paintings filtered by artists
¤ But how we filter the Artists facet values by nationality,
gender, age, etc.?
¨ Exploring contents at level of sets using semantic
relationships, e.g.
¤ The museums that have bronze Greek statues
¤ “Women portrayed by women”: paintings with subject:woman
and artist:gender:female
¤ Schools attended by the daughters of U.S. democratic
presidents (http://www.freebase.com/labs/parallax/)
¤ Challenges: effective models and usable interface
¨ An example: Sewelis
46. Beyond binary classication | 1
¤ Classification (faceted or not) is usually
binary:
¤ An item must be either relevant (1) or
not relevant (0) to a certain category
¤ Problem: quite arbitrary decision in
many real domains
47. Beyond binary classication | 2
î How to classify acathedral by architectural
style?
¤ Built upon a 6th century buliding
¤ Mainly gothic
¤ 17th century (baroque) towers
¤ Rebuilt during neoclassicism
¤ Decorations added in 19th century
¤ Contains Roman forum marbles (donated by Pius
IX)
¤ …
î Do we tag the cathedral with all or only some of
these?
î A classification may be correct for a kind of users but
ineffective for another one
48. Beyond binary classication | 3
î Monna Lisa is a well known
portait of a woman, but…
î There is also a landscape in
the background
î Do we classifity it as
“subject: woman” and
“subject: Tuscan landscape”
too?
49. Beyond binary classication | 4
î Onion is very
used in French
cuisine
î How do we
distinguish
“onion-based”
recipes from all
the recipes with
onion inside?
50. Beyond binary classication | 5
¨ A possible solution:
associating weights
to each triple item-
facet-value
¤ A statement about
the statement
¨ Values between 0 and
1 or other scales
¨ Query could be
specified in terms of
facet-values pairs and
ranges of weights
51. Beyond binary classication | 6
¨ Subjective weights
¤ Relevance: at which
extent the item can be
considered as belonging
to a certain facet value
¤ Significance: the relative
importance of the item
according to a facet value
¨ Objective weights
¤ E.g. Concentration or quantity (e.g.
a thing is made for the 10% of
material:bronze)
¤ E.g. for exploring venues:
distance from points of interests
53. Handling information overload
53
¨ Too more facets and facets values may
generate information overload too!
¤ Possible solution: Display only the most relevant
facets (and facet values) for the user profile or the
given context
¨ How to determine the most “interesting” facets in
a given context?
¤ E.g. those with a less “uniform” distribution of
values (more correlation)
¤ We will discuss this in a next lecture… :-)
54. Interested in MS Theses? Contact us! :-)
54
¨ Advisors: Prof. Di Blas, Prof. Paolini
¨ Both theoretical and development
¨ Fuzzy facets
¨ Semantic faceted search
¨ Advanced visualizations
¨ …
¨ Your own ideas! :-)
55. 55
Any final questions?
Are you still alive/awake?
Thank you for your attention!