Personalization for E-Commerce

Transcription

Personalization for E-Commerce
Personalization for E-Commerce
Tutorial presented at UM 2001
Anthony Jameson
German Research Center for
Artificial Intelligence (DFKI)
Saarbrücken, Germany
jameson@dfki.de
http://dfki.de/~jameson
Table of Contents
3
Introduction
What Is Personalization in General?
Types of Systems Covered
Emphasis in This Tutorial
Availability of These Slides
Recommending Products
PersonaLogic
PurchaseSource
The FindMe Systems
Casper
Lifestyle Finder
Amazon.com Recommendations
Comparison
Facilitating Navigation
Clixsmart Navigator
Adaptive Information Server
Adaptive Web Sites
Angara Converter
7d b-bot
Comparison
Allowing Configuration
MyYahoo!
InfoBeans
Comparison
Using Lifelike Characters
The Notebook Expert
Artificial Life
The PPP Persona
The MIAU Agents
REA
Comparison
5
5
6
7
8
9
9
12
24
34
38
41
45
53
53
55
62
67
75
84
87
87
99
107
109
109
120
130
137
146
152
5
Introduction
6
Introduction
5
What Is Personalization in General?
Adaptability
Adaptation
Anthropomorphism
Personalization can be seen as the union of three overlapping system
properties
Types of Systems Covered
6
1. Product recommendation systems
2. Systems that guide users to the appropriate pages in a hypertext
system
3. Systems that users can configure to suit their own requirements and
preferences
4. Lifelike characters
7
Introduction
8
7
Emphasis in This Tutorial
Technology for
personalization
Personalized
systems
Enhanced
user
experience
Increases in:
Customer acquisition
Cross−selling
Up−selling
Customer loyalty
...
Availability of These Slides
8
• Several companies supplied information for this tutorial on the
condition that it would be made available only to the tutorial
participants
• Accordingly, an electronic version of these slides cannot be made
available via the World-Wide Web
• For the same reason, the printed slides may be photocopied only on
a small scale (e.g., for individual colleagues)
• Those who require electronic versions of particular parts of the
slides can contact the presenter or the companies in question
9
Recommending Products / PersonaLogic
10
Recommending Products
http://www.personalogic.com. As of July, 2001, the system is no longer
available at this URL. Specific parts of it are available − for example, at
http://www.purina.personalogic.com, from which the following examples are
taken.
9
Example Screens (1)
At PersonaLogic, U can obtain recommendations about a broad range
of "products" (including colleges, dogs, and political candidates)
To get recommendations concerning dogs, U first answers questions
about her evaluation criteria
Example Screens (2)
Whenever U visits the "YOUR RESULTS" page, S presents the
products that score best according to the evaluation criteria that U has
specified
10
11
Recommending Products / PersonaLogic
Multi-Attribute Utility Theory (MAUT)
11
For general treatments of MAUT, see von Winterfeldt, D., & Edwards, W.
(1986). Decision analysis and behavioral research. Cambridge, England:
Cambridge University Press.; Clemen, R. T. (1996). Making hard decisions:
An introduction to decision analysis. Pacific Grove, CA: Duxbury.
12
Role in personalized product presentation
• Many (though not all) systems for personalized product presentation
use some variant of MAUT
• The particular way in which it is used can vary greatly
Basic principles
• Each object O has a number of attributes that are relevant to its
evaluation by U
• Each attribute has an importance weight for U
• U could in principle assign to each attribute of O a value (e.g., on a
scale from 0 to 10)
O to U is the sum of the values of its attributes,
weighted by the importances of the attributes
• The overall value of
• What
•
U wants is the object O with the highest overall value
Note: "Price" is usually viewed as an attribute, usually one with a
high importance weight
PurchaseSource
http://www.frictionless.com/. As of July, 2001, the demo that was used to
make the following slides is no longer available.
Introduction
• Frictionless.com created a web-based system (S) called
PurchaseSource
• On-line vendors can use this system to help customers locate
suitable products
• What we will see on the following slides is a walkthrough of a
demonstration of this system
• What is of interest is the basic system itself, not the specific
information that has been entered on the particular type of product
that our hypothetical U is looking for (backpacks)
12
13
13
Recommending Products / PurchaseSource
14
Overview of Types of Sports Articles
•
U has
already
indicated that
he is
interested in
sports articles
•
U now clicks
on
"Backpacks"
under
"Outdoors"
Choosing a "Profile"
14
• By choosing the
profile "Day Hiker",
U allows S to make
guesses about
many of U’s
preferences
•
U will now need to
specify only the
preferences that S
has not guessed
correctly
15
15
Recommending Products / PurchaseSource
16
Specifying Preferences
• For each attribute,
U can
specify how important it is
that the product should fall
into the specified range
•
U can learn more about the
attribute by clicking the
"learn more" button
•
•
For the attribute
"Capacity", see the result
on the slide after next
U can also "edit" the range
itself
•
Here, our U chooses to
edit the range for
"Capacity" by clicking on
the first "Edit" button
Editing the Range for an Attribute
• Before editing the
16
range, our U
chooses to "learn
more" about the
attribute "Capacity"
17
17
Recommending Products / PurchaseSource
18
Learning More About an Attribute
• After reading this
information, our U
goes "back" to the
previous screen
U uses the
right-hand pop-up
menu to change the
maximum capacity
to 2,500 cu. in
• There,
19
19
Recommending Products / PurchaseSource
20
Overview of the Recommended Products (1)
• Each recommended
product is
summarized here in
a single line
Overview of the Recommended Products (2)
• Our
20
U clicks on the
light-colored bar for
the last product
listed ("Jansport
Couloir") to find out
more about the
reasons for the
recommendation
21
21
Recommending Products / PurchaseSource
22
Summary for a Single Product
• This page refers to a
single product, but it still
doesn’t include all of the
detailed considerations
underlying S’s rating of
this product
U wants to know
these details, so he clicks
on "Details"
• Our
Details for a Single Product
22
U can see, for
each attribute, how S
• Here,
rated this product with
respect to this attribute
• (Only the first part of the
long screen is shown
here)
23
23
Recommending Products / PurchaseSource
24
Comparison of Three Products
• Later, our
U chooses to request
a comparison of three other
backpacks, which were more
highly rated
• In each column of this screen,
essentially the same information
is presented as in the screen
with details for a single product
(see the previous slide)
The FindMe Systems
Burke, R. D., Hammond, K. J., & Young, B. C. (1997). The FindMe approach
to assisted browsing. IEEE Expert, 12(4), 32−40.
Burke, R. D. (2000). Knowledge-based recommender systems.
Encyclopedia of Library and Information Science.
http://www.ics.uci.edu/~burke/research/cbr-ec.html
Introduction
24
The FindMe family
• Burke and his colleagues developed several related systems that
are based on the "Find-Me" approach to recommendation
• The basic principles are illustrated here with the restaurant
recommendation system "Entree", which is available on the web
Overview of Entree
Specification of preferences
• At the beginning of a session,
U has two options:
1. U specifies her preferences by choosing a few properties of
restaurants
2. U enters the name of some other restaurant that she knows (in
Chicago or another city) which exemplifies her preferences
Criticism of the initial results
• S presents some restaurants that correspond to
U’s specifications
• U can then specify how the set of candidates can be improved
25
Recommending Products / The FindMe Systems
The Entree-System is available under http://infolab.ils.nwu.edu/entree/pub/
25
26
Entree: Examples (1)
U specifies her interests
Entree: Examples (2)
S offers an initial result, further results, and an opportunity to
criticize the results
26
27
27
Recommending Products / The FindMe Systems
28
Entree: Examples (3)
U would prefer a similar restaurant with
South American cuisine
Entree: Examples (4)
28
Faced with this result, U clicks on the button "Creative" in order to find a
similar but more creative restaurant
29
Recommending Products / The FindMe Systems
30
Entree: Examples (5)
29
By now the set of candidates has been reduced to a single restaurant
Key Methods (1)
30
This example illustrates some of the methods that are applied in the
FindMe approach:
Retrieval by similarity
1. At first, U specifies either
• An example product; or
• A small set of attributes that correspond to various goals
•
Examples: "cuisine" = "French", "style" = "casual"
2. S orders all of the products in the database according to their
similarity with the given example or attribute set
• For each goal, a similarity metric has to be defined
• For example, S needs to know how similar French and Japanese
cuisine are
• Defining these similarity metrics is one of the main tasks that has
to be handled when a system like Entree is introduced in a new
domain
31
Recommending Products / The FindMe Systems
32
Key Methods (2)
31
Different orderings of goals
• When restaurants are sorted, the various goals are taken into
account in a particular order
• Example:
•
First they are sorted in terms of the similarity of their cuisine to
U’s specification
•
Then restaurants that are equally similar with respect to cuisine
are sorted with respect to "atmosphere"
•
Assumption here: cuisine is a more important criterion than
atmosphere
• Handling differences among users in the importance of goals
•
•
In some of the FindMe systems, U can specify the relative
importance of the goals herself
She can choose from several prespecified "retrieval strategies"
(e.g., "Money is no object")
Key Methods (3)
32
Criticizing (tweaking) of the candidate restaurants by U
U, S simply removes from the candidate set all
restaurants that do not satisfy the specified constraint
• After a critique by
•
Example: All restaurant that are not "more creative" than the most
recently suggested one
Use of familiar concepts
•
S makes it possible for U to employ concepts that are more familiar
and less concrete than the attributes stored in the database (e.g.,
casual)
33
Recommending Products / The FindMe Systems
34
Key Methods (4)
33
Explaining conflicts among specifications
•
U may sometimes specify a combination of desired attributes that
(almost) never occurs
•
Example from the automobile domain: powerful motor combined
with low gas consumption
• In this case
•
S calls attention to the conflict
Note: This functionality is not realized in Entree, but it is realized
in the related system Car Navigator (Burke et al., 1997)
Casper
Bradley, K., Rafter, R., & Smyth, B. (2000). Case-based user profiling for
content personalisation. In P. Brusilovsky, O. Stock, & C. Strapparava
(Eds.), Adaptive hypermedia and adaptive web-based systems: Proceedings
of AH 2000 (pp. 62−72).Berlin: Springer.
Case-Based Reasoning (1)
34
Casper: Personalization in a job finding system
• Underlying retrieval system: JobFinder, an Irish recruitment web site
• JobFinder provides conventional search-engine functionality over a
database of job offers
• Limitation:
•
U’s search query seldom reflects all of her actual criteria
U ratings of jobs previously retrieved by JobFinder represent an
additional source of information on U’s needs
• These previous "cases" are used to filter further cases retrieved by
the system
35
Figure 1 of Bradley et al. (2000)
35
Recommending Products / Casper
Case-Based Reasoning (2)
Overview of the two-stage retrieval process in Casper
Case-Based Reasoning (3)
Figure 3 of Bradley et al. (2000)
36
S’s criteria for assessing the similarity between two cases takes into
account the domain-specific significance of the various attributes
36
37
Recommending Products / Casper
38
Case-Based Reasoning (4)
37
B
G
G
G
B
B
?
B
G
G
B
B Bad
G
G Good Jobs
? Candidate Job
Figure 4 of Bradley et al. (2000)
Procedure
• The relevance of a new job offer is predicted on the basis of the
ratings of the k (here: 5) most similar previous cases ("nearest
neighbors")
• Each of these previous ratings is weighted by the similarity of the
case to the current job offer
Evaluation results (summary)
• About 70% correct classifications, using for each
57 job offers previously rated by that U
U a case base of
Lifestyle Finder
The Lifestyle Finder was freely available on the web long enough to collect
data from 20,000 users, but it is no longer available. Further information is
given in the section "Inference: Data-Based". This screen shot is Figure 3 of:
Krulwich, B. (1997). Lifestyle Finder: Intelligent user profiling using
large-scale demographic data. AI Magazine, 18(2), 37−45.
Elicitation of Demographic Information
38
The Lifestyle Finder elicits demographic information in a playful fashion
that does not require U to supply identifying information
93% of the users surveyed agreed that the Lifestyle Finder’s questions
did not invade their privacy
39
The system recommends 15 web pages, of which 3 are chosen at random
for evaluation purposes. This slide shows part of Figure 4 of: Krulwich, B.
(1997). Lifestyle Finder: Intelligent user profiling using large-scale
demographic data. AI Magazine, 18(2), 37−45.
39
Recommending Products / Lifestyle Finder
40
Demographically Based Recommendation (1)
After U has answered a few lifestyle-related questions (see the section
"Properties"), the Lifestyle Finder’s presents its guess about the
demographic cluster to which U belongs
It then recommends web pages with information about products that
should be of interest to persons within this cluster
Demographically Based Recommendation (2)
40
Procedure of "demographic generalization"
• The Lifestyle Finder assigns U to one or more of 62 demographic
clusters from a commercial demographic system employed for
consumer marketing
• If U’s answers match more than one cluster, the predictions that are
common to all matched clusters form the basis for recommendations
for U
• S may ask further questions that efficiently narrow down the set of
possible clusters for U
Evaluation
• Users gave positive responses to more than 40% of the web pages
recommended in this way, compared with about 30% of the
randomly recommended pages
• This level of accuracy is lower than that attainable through other
methods
• This method is accordingly proposed as a complement to methods
that require more input from each user
41 Recommending Products / Amazon.com Recommendations 42
Amazon.com Recommendations
Introduction
41
• Amazon.com is a pioneer in the area of commercial product
recommendation
• The recommendations are viewed as an additional service which is
intended to stimulate further purchases once U has purchased one
or more books
• As we will see, the techniques employed are less well suited for
cases in which a customer who has specific requirements asks for
recommendations
https://www.amazon.com/
Examples (1)
A link to the recommendation page is present in the introductory page
42
43 Recommending Products / Amazon.com Recommendations 44
43
Examples (2)
U sees this page once she has clicked on "See more Books
recommendations"
Examples (3)
44
On this page, U can influence the recommendations by explicitly rating
the products she has bought so far (here: only 1 product)
45
Recommending Products / Comparison
46
Comparison
45
PersonaLogic
Comparison of Approaches (1)
PurchaseSource
FindMe
Casper
Lifestyle Finder Amazon.com
1. What sort of explicit input does U have to supply?
Importances of General buyer
attributes
category;
Importances
and desired
levels of
attributes
Example or
Normal queries;
features of
Ratings of
desired product; offers
Brief critiques of
suggested
products
Answers to
None; or ratings
amusing
of products
questions about
demographic
characteristics
2. What is the minimum information that U must supply before receiving her first tailored
recommendation?
Importances for General buyer
category
one or more
attributes
Normal query
Example or
features of
desired product
Answers to the Purchase or
questions
rating of at least
one product
Comparison of Approaches (2)
PersonaLogic
PurchaseSource
FindMe
Casper
46
Lifestyle Finder Amazon.com
3. To what extent (and how) does S ensure that U has enough general knowledge about
the type of product in question to be able to make a well-founded decision?
Explanatory
texts for
individual
attributes
Explanatory
texts for
individual
attributes
Use of familiar
concepts;
sometimes
explanation of
conflicts
Not at all
Not at all
Some
informative
texts on site
4. How expressive are the means that U has available for communicating her
requirements to S?
Very low
Indirect and
Extremely high Limited
Quite high
expressiveness, expressiveness, expressiveness expressiveness coarse
within MAUT
within MAUT
framework
framework
Low
expressiveness
47
47
PersonaLogic
Recommending Products / Comparison
48
Comparison of Approaches (3)
PurchaseSource
FindMe
Casper
Lifestyle Finder Amazon.com
5. What properties of U are modeled?
Importance
weights of
attributes
Importance
weights and
desired values
of attributes
Desired
properties of
product;
sometimes
relative
importance of
attributes
Membership in No modeling of
Ratings of
previous offers a demographic user properties
group
(no explicit
modeling)
6. What procedure does S use to arrive at recommendations?
MAUT
MAUT
Case-based
Mainly
computation of reasoning
similarities
Demographic
marketing
research
conducted
beforehand
Collaborative
filtering
Comparison of Approaches (4)
PersonaLogic
PurchaseSource
FindMe
Casper
48
Lifestyle Finder Amazon.com
7. In what form does S give its recommendations?
Presentations
Overall
evaluation and at different
levels of detail
description
Brief textual
description
Ordering of
offers
Briefly
annotated links
to external web
pages
Brief
description,
perhaps
supplemented
with reviews
8. To what extent can U see that a recommended product is better for her than particular
other products?
Tabular
comparisons
Little support
Tabular
comparisons at
different levels
of detail
Little support
Little support
Little support
49
49
PersonaLogic
Recommending Products / Comparison
50
Comparison of Approaches (5)
PurchaseSource
FindMe
Casper
Lifestyle Finder Amazon.com
9. Does S adapt the way in which it presents individual products to U?
No
Presentation
refers to U’s
requirements
No
No
No
No
10. To what extent can U revise the specification of her requirements after seeing some
recommendations?
Possible but not Possible but not Specifically
well supported well supported supported
Through
Virtually
revised queries impossible
To limited
extent
Comparison of Approaches (6)
PersonaLogic
PurchaseSource
FindMe
Casper
50
Lifestyle Finder Amazon.com
11. To what extent does the recommendation process correspond with U’s natural way
of thinking?
Well
Well
Well
Well
Poorly
Well
12. To what extent can U understand the reasons why S made particular
recommendations?
Little support
Good support
Little support
No support
No support
No support
51
51
PersonaLogic
Recommending Products / Comparison
52
Comparison of Approaches (7)
PurchaseSource
FindMe
Casper
Lifestyle Finder Amazon.com
13. To what extent does S take into account the fact that the interests of U can vary from
one situation and time to the next?
Not at all
Problem cannot Problem cannot Problem cannot Most recent
arise
arise
arise
cases could be
given greater
weight
Only gradual
adaptation
possible
14. To what extent can S reuse what it has learned about U in connection with a
particular recommendation process later, when making recommendations to U in
another situation?
Currently not
supported, in
principle
possible
Currently not
supported, in
principle
possible
Currently not
supported, in
principle
possible
Well
Well
Very well, for
products of the
same type
Comparison of Approaches (8)
PersonaLogic
PurchaseSource
FindMe
Casper
Lifestyle Finder Amazon.com
15. When S is set up for use with a new type of product, to what extent do the designers
have to make use of knowledge about that type of product?
A great deal
A great deal
A great deal
A good deal
To some extent Hardly any
52
53
Facilitating Navigation / Clixsmart Navigator
54
Facilitating Navigation
Smyth, B. (2001). Clixsmart navigator: A briefing for mobile and wireless
portal operators. Dublin: ChangingWorlds. White paper.
53
Reducing Click Distances: Before
To access the movie listings of her local theater via a WAP portal, the U
in this example must make 29 clicks, often scrolling in the process
Reducing Click Distances: After
After S has adapted the menu structure to U’s access patterns, only 2
clicks are required
54
55
Facilitating Navigation / Adaptive Information Server
56
Adaptive Information Server
Background
The demos shown during the tutorial are available from the company’s web
site: http://www.adaptiveinfo.com.
55
• The firm AdaptiveInfo was founded recently as a spin-off from the
University of California at Irvine
• It’s technology is based largely on research by Daniel Billsus and
Michael Pazzani
• Representative articles:
•
Billsus, D., & Pazzani, M. J. (1999). A hybrid user model for news
story classification. In J. Kay (Ed.), UM99, User modeling:
Proceedings of the Seventh International Conference
(pp. 99−108).Vienna: Springer Wien New York.
http://www.ics.uci.edu/~pazzani/Publications/Publications.html
•
Billsus, D., & Pazzani, M. J. (2000). User modeling for adaptive
news access. User Modeling and User-Adapted Interaction, 10,
147−180.
The slides in this section were adapted from slides supplied by Michael J.
Pazzani.
Adaptive Information Server: Overview
• Automatically learns a profile of user interests through normal
interaction
• Prioritizes presentation of content and commerce opportunities
• Proven to increase page views by over 40%
• Designed for scalable data mining
•
Real time: personalization
•
Aggregated for marketing insight
• Integrates with existing wireless platforms
56
57
Facilitating Navigation / Adaptive Information Server
Adaptive News Server
57
Some of the technology underlying the Adaptive Information Server is
described by: Billsus, D., & Pazzani, M. J. (2000). User modeling for
adaptive news access. User Modeling and User-Adapted Interaction, 10,
147−180.
58
A
B
C
The Adaptive News Server spontaneously adapts its selection of news
items on the basis of the user’s reading behavior
Adaptive Classified Server
58
The good news:
• Individual classified ads fit on most cell
phone screens
The bad news:
• There are 250,000 ads in a paper, but
users won´t scan through 25 or press 25
keys to describe what they want
• A newspaper costs $0.25
• The classified ad server must provide
more benefits than a newspaper
59
59
Facilitating Navigation / Adaptive Information Server
60
Before Personalization
After Personalization
60
61
61
Facilitating Navigation / Adaptive Information Server
62
Extra Functions With Phones
Adaptive Web Sites
Perkowitz, M., & Etzioni, O. (2000). Towards adaptive web sites: Conceptual
framework and case study. Artificial Intelligence, 118, 245−275.
Introduction and Overview (1)
62
Background
• The "adaptive web sites" approach is being developed by Mike
Perkowitz and Oren Etzioni at the University of Washington
• The approach has attracted a good deal of attention, although
apparently so far only a prototype implementation exists
Overview of the method
• The system analyses logs of the use of a given web site
•
S proposes new index pages, each of which contains a number of
related links that are likely to be of interest to the same type of user
63
Facilitating Navigation / Adaptive Web Sites
64
Introduction and Overview (2)
63
4
Music Machines
2
Music Machines
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
Music Machines
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
Music Machines
Music Machines
Music
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
Machines
5. Linn LinnDrum samples
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
Music Machines
Music Machines
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
Music Machines
1. Boss Dr-110 samples
2. Boss Dr-55 samples
3. Korg KPR-77 samples
4. Korg KR-55 samples
5. Linn LinnDrum samples
1
3
5
Introduction and Overview (3)
• IndexFinder generates a candidate index page
• The human web master accepts or rejects the candidate page
• If the page is accepted, IndexFinder creates a final version of the
page and adds it to the web site
• The web master specifies
•
the page’s title
•
where in the site it should be linked
• Rejected pages are discarded
64
65
65
Facilitating Navigation / Adaptive Web Sites
66
The Basic Data
Figure 1. Typical user access logs, these from a computer science Web site. Each entry corresponds
to a single request to the server and includes originating machine, time, and URL requested. Note
the series of accesses from each of two users (one from SFSU, one from UMN).
24hrlab-214.sfsu.edu - - [21/Nov/1996:00:01:05 -0800] "GET /home/jones/collectors.html HTTP/1.0" 200 13119
24hrlab-214.sfsu.edu - - [21/Now/1996:00:01:06 -0800] "GET /home/jones/madewithmac.gif HTTP/1.0" 200 855
24hrlab-214.sfsu.edu - - [21/Nov/1996:00:01:06 -0800] "GET /home/jones/gustop2.gif HTTP/1.0" 200 25460
x67-122.ejack.umn.edu - - [21/Nov/1996:00:01:08 -0800] "GET /home/rich/aircrafts.html HTTP/1.0" 404 617
x67-122.ejack.umn.edu - - [21/Nov/1996:00:01:08 -0800] "GET /general/info.gif HTTP/1.0" 200 331
203.147.0.10 - - [21/Nov/1996:00:01:09 -0800] "GET /home/smith/kitty.html HTTP/1.0" 200 5160
24hrlab-214.sfsu.edu - - [21/Nov/1996:00:01:10 -0800] "GET /home/jones/thumbnails/awing-bo.gif HTTP/1.0" 200 5117
• Typical user access logs, these from a computer science web site
• Each entry corresponds to a single request to the server
• It includes the originating machine, the time, and the URL requested
• Note the series of accesses from each of two users (one from
SFSU, one from UMN).
Inner Workings of IndexFinder
66
IndexFinder consists of three basic modules:
1. The log processing module takes the web server access logs and
computes how often pages co-occur in user visits
2. The cluster mining module takes this information and a graph of the
web site and finds clusters of frequently co-occurring pages
3. The conceptual clustering module uses conceptual descriptions of
the site s pages to convert these clusters into coherent concepts
• These are output as candidate index pages and presented to the
web master.
67
Facilitating Navigation / Angara Converter
68
Angara Converter
Experience of the Unknown Visitor (1)
The slides in this section are based on the "tour" offered at
http://www.angara.com.
67
Experience of the Unknown Visitor (2)
68
• The Fashion Store highlights its most frequently purchased apparel
•
•
•
•
•
•
•
item on the home page, a men’s leather jacket
Although this week’s featured item effectively draws men into the
site, women must click through to specific departments before
finding merchandise of personal interest
This means conversion rates are not optimized
The Fashion Store is experiencing the same phenomenon as most
of the on-line industry:
• The majority of browsers visit a home page and leave before
going deeper into the site, or purchase product, because the
home page information is targeted to only one audience
Increasing the conversion rates for browsers to buyers will require
targeted, relevant content
Over 90% of visitors to Web sites are unknown
Industry conversion rates for new visitors average between 1 and
2%
Site abandonment rates are 85% at the home page
69
69
Facilitating Navigation / Angara Converter
70
Targeting Content: The Process (1)
Targeting Content: The Process (2)
70
• The Fashion Store came to Angara for help to develop a strategy to
increase their conversion rates while maintaining their uncluttered
home page
• We worked with them to understand key business drivers and,
drawing on extensive segmentation expertise, helped develop a
segmentation scheme for The Fashion Store’s site that ultimately
raised conversion rates two-fold
71
71
Facilitating Navigation / Angara Converter
72
Targeting Content: The Results
Now, unknown visitors to the Fashion Store site are presented a
customized home page presenting targeted "Feature Items" based on
their prospect category − defined by geographic and demographic
variables such as age and gender.
Measurement and Reporting of Results (1)
72
73
73
Facilitating Navigation / Angara Converter
74
Measurement and Reporting of Results (2)
• Real-time access to reporting through a secure Website provides
The Fashion Store with detailed information to illustrate the
improvement in conversion rates of unknown visitors based on
targeting
• These reports provide the information necessary for The Fashion
Store to quickly react to trends in customer behavior
• Custom content can be frequently modified at The Fashion Store
site in order to surface new offers, react to inventory changes, or
address other changes in market conditions
Calculating Return on Investment
74
75
Facilitating Navigation / 7d b-bot
76
7d b-bot
Introduction
Company URL: http://www.7d.net. The graphics in the following slides were
supplied by Christoph Hölscher.
75
Background
• 7d is a relatively new company, located in Hamburg, which is now
introducing its first products
The "b-bot"
• We will look mainly at the system b-bot
• The main function of this product − at least initially − is the
recommendation of informative web pages within large web sites
Overview of Recommendation Approaches (1) 76
• One of the distinguishing characteristics of 7d’s approach is shown
in the following two graphics
• Three different recommendation techniques are integrated, which
are usually found in different systems:
1. Content-based recommendation
2. Behavior-based recommendation
3. Rule-based recommendation
77
77
Facilitating Navigation / 7d b-bot
78
Overview of Recommendation Approaches (2)
Content-Based Recommendation
Goal: Given a particular document, recommend similar documents
78
79
79
Facilitating Navigation / 7d b-bot
80
Behavior-Based Recommendation
• In the example,
U is assigned to the second (red) group on the
basis of his page selections to date
• Accordingly, S recommends pages that members of this group have
visited and/or rated positively
• Information about users consists mainly of page selections (as
opposed to explicit ratings)
URL of ATG: http://www.atg.com/
Rule-Based Recommendation
This graphic indicates how rules are defined and applied in the b-bot
80
81
ATG’s Dynamo Personalization Server and two other personalization servers
are analyzed in depth by: Fink, J., & Kobsa, A. (2000). A review and analysis
of commercial user modeling servers for personalization on the world wide
web. User Modeling and User-Adapted Interaction, 10, 209−249.
81
Facilitating Navigation / 7d b-bot
82
Related Examples From ATG (1)
Even without technical knowledge, an administrator can define user
groups in terms of particular attributes
Related Examples From ATG (2)
In a similar way, the administrator can specify which objects within a
given category should be shown to which groups
82
83
83
Facilitating Navigation / 7d b-bot
84
Support for Analysis and Evaluation
As with many other personalization servers, a monitoring component
allows various types of analysis of user’s behavior
Comparison
Comparison of Approaches (1)
Clixsmart Navigator
Adaptive Information
Server
Adaptive Web Sites
Angara Converter
84
7d b-bot
1. Are adapations made to individual users or to groups?
Individuals
Individuals
To date no tailoring;
Groups
development of
tailoring for subgroups
is planned
Individuals and/or
groups
Generation of new
index pages
Changes in content,
layout, and/or
structure
2. What form do adaptations take?
Rearrangement of
menu hierarchies
Rearrangement of
options within menus
Changes in content of
pages
3. To what extent do the adaptations demand U’s attention?
Unexpected changes
might be distracting
Unexpected changes
might be distracting
No attention
demanded
U may never notice
adaptation
(Depends on
particular form of
adaptation)
85
85
Clixsmart Navigator
Facilitating Navigation / Comparison
86
Comparison of Approaches (2)
Adaptive Information
Server
Adaptive Web Sites
Angara Converter
7d b-bot
4. What sort of explicit input is U required to supply?
None
None
No explicit input (later Some specification of
version: perhaps a bit) demographic
information
In some cases:
ratings, specification
of demographic
information
5. How much additional effort is required to administer the system?
Little or no
domain-specific
administration
Little or no
domain-specific
administration
Collection and use of
Webmaster
postprocesses
marketing data
proposed index pages
(Depends on parts of
system used)
87
Allowing Configuration / MyYahoo!
88
Allowing Configuration
Background
87
• Yahoo! was one of the first web portals to employ personalization
methods
• MyYahoo!, introduced in July, 1996, offers a rich variety of
personalization options, which have been used by millions
• The following slides
http://www.yahoo.com
1. illustrate some of the ways in which MyYahoo! supports
configuration by the user
2. summarize some lessons learned over the years by the
designers of MyYahoo!
A User’s Main Page
88
89
89
Allowing Configuration / MyYahoo!
90
Changing the Layout of a Page
Changing the Content of a Page
90
91
91
Allowing Configuration / MyYahoo!
92
Changing the Content of a Module
Adding a Page With One Click
92
After reaching an interesting page that is not currently included in his
main page, the user can have it included simply by clicking on the "Add
to MyYahoo!" button at the upper right
93
93
Allowing Configuration / MyYahoo!
94
Creating New Pages
More ambitious users can create additional pages with collections of
links
The remaining slides in this section contain abbreviated excerpts from the
following article: Manber, U., Patel, A., & Robison, J. (2000). Experience with
personalization on Yahoo! Communications of the ACM, 43(8), 35−39.
Configuration by the User
94
Personalization within individual modules
• For example, users can choose which TV channels they want to
include in their TV guide in addition to which local cable system they
use
Simplification of configuration actions
• Modules can be selected from a (long) list, but can also be added by
clicking on a button at the original content page
• For example, every weather page (http://weather.yahoo.com)
contains an "Add to My Yahoo!" button, which adds that page
directly to the user’s My Yahoo! page
• Also, each module on a My Yahoo! page has an edit and a remove
button, allowing users to manipulate their pages directly, without
ever needing to visit an edit/layout screen
95
Allowing Configuration / MyYahoo!
96
Adaptation by the System
95
Simple automatic adaptation of content
• Example: a sports module that lists the teams in the user’s area
after obtaining that information from U’s profile
Adaptation of search results
• In some cases we can complement the usual Web search with
direct, focused content that can sometimes be personalized
• For example, if
U searches for the name of a current movie, we
point to Yahoo! movies, show an image from the movie, the cast,
and a pointer to a page with current showtimes
• If
U looked at the showtimes page previously and entered a zip
code, that page is automatically customized to show the movie
theaters near U
• With one click after searching for a movie name,
showtimes in U’s area
U can see the
Predictability as a Paramount Goal
96
• Most users expect to have at least an intuitive notion of what is
given to them, and they expect to see the same behavior
consistently
• Being surprised is wonderful if it is entirely a positive surprise, but
overall, being unpredictable is a negative
• In particular, if people are not sure how things work, they are less
prone to experimentation, because they are afraid of breaking
something, or getting into a state that cannot be undone
• In the case of news, for example, it is not clear that people want
personal news; they often want the same news everyone else is
getting
• Getting local weather and news about a local sports team from zip
codes is obvious
• Getting news about cancer because the user read some medical
journals in the past or searched for some medical terms can confuse
the user at best, and at worst, can jeopardize user trust and raise
serious privacy concerns in the user’s mind
97
97
Allowing Configuration / MyYahoo!
Other Lessons From MyYahoo!
• Most users take what is given to them and never customize
• A great deal of effort should go into the default page
• Power users will do amazing things; never underestimate them
• People generally don’t understand the concept of customization
98
99
Allowing Configuration / InfoBeans
100
InfoBeans
Background
This and the following slides include abbreviated excerpts from: Bauer, M., &
Dengler, D. (1999). InfoBeans: Configuration of personalized information
assistants. In M. T. Maybury (Ed.), IUI99: International Conference on
Intelligent User Interfaces. New York: ACM.
99
Origins
• InfoBeans was developed as a research prototype at DFKI
• It offers users novel ways of configuring web pages
• Some of its most important properties are illustrated in the example
scenario shown on the following slides
Concepts
• An InfoBean is a configurable component that either
•
•
encapsulates an existing information service; or
specifies a process of information gathering and providing
• It communicates with its environment through channels which pass
information to or obtain data from other components
• A collection of InfoBeans with a common user interface (a WWW
document accessible by a standard Web browser) is called an
InfoBox
Example Scenario
•
100
U is a business traveler who plans all her trips using a web-based
interface to an international flight and hotel reservation system
• She wants to have access to
1. detailed hotel information
2. a city map showing the location of the hotel
3. local weather data
• She has found excellent web-based services for each of these
features
• Since similar information is required several times, she wants to be
able to access only one page that provides her all relevant data
quickly and automatically
101
Allowing Configuration / InfoBeans
102
Example of an InfoBox
101
Use of the InfoBox
• The figure shows the resulting web page
• When
U now submits new hotel request information, the two
InfoBeans considering weather and hotel information are
immediately activated
• The city map InfoBean starts working after the hotel information
InfoBean has passed the street name into its output channel
102
103
Allowing Configuration / InfoBeans
Configuration Procedure (1)
103
•
104
U loads the InfoBox start page and has access to
1. predefined InfoBeans
2. her own earlier specified InfoBoxes
3. InfoBeans and the option to add or change things
•
U decides to add an InfoBox called "Business Travel"
•
•
Her browser jumps to a new nearly empty page containing only
the InfoBean toolbar
U defines a new InfoBean to encapsulate her usual travel
reservation service
•
U is asked to draw a region on the still empty InfoBox page that
should serve as the display area of this InfoBean
•
U specifies it as a pure encapsulation:
•
She simply enters the URL of the reservation service
Configuration Procedure (2)
104
• The top left frame is the HTML form for requesting hotel information
• The relevant data to be caught for delivery with respect to a current
stay are
1. the name of the city
2. the hotel
3. the date
• An InfoBean provides the ability to specify output channels easily via
direct manipulation
• In the case of defining the output channel City this means that the
user selects the text input field named ’City’
• The InfoBean compiles this selection into the appropriate wrapper
action in order to gather the specified information for delivery from
the document
Color profile: Generic CMYK printer profile
Composite Default screen
105
Figure 5.7 from: Bauer, M., & Paul, G. (2001). Programming by
demonstration for information agents. In H. Lieberman (Ed.), Your wish is my
command: Programming by example (pp. 87−114). San Francisco: Morgan
Kaufmann.
105
Allowing Configuration / InfoBeans
106
Specifying
More Complex Wrappers (1)
Figure 5.7
An infobox
like the one shown in the video
A sample InfoBox.
Specifying More Complex Wrappers (2)
106
• If time permits, a video will be shown that illustrates how a userS can
specify a relatively difficult wrapper
V:\002564\002564.VP
Monday, December 18, 2000 12:42:22 PM
R
L
107
Allowing Configuration / Comparison
108
Comparison
107
Comparison of Approaches
MyYahoo!
InfoBeans
1. To what extent does U configure S?
To a great extent, in part with especially
convenient methods
To a great extent, with intelligent support
2. To what extent does S configure itself spontaneously?
In carefully selected cases
S interprets U’s specifications in an intelligent
way
109
Using Lifelike Characters / The Notebook Expert
110
Using Lifelike Characters
Self-Description of Soliloquy (1)
See the Corporate Fact Sheet, which is available via
http://www.soliloquy.com. A similarly upbeat description of the firm is given in
the following article: Lucente, M. (2000). Conversational interfaces for
e-commerce applications. Communications of the ACM, 43(9), 59−61.
109
From the Corporate Fact Sheet
"Soliloquy’s Dialogue Experts allow shoppers to interact with a
website in the most natural and intuitive way: by using their own
words"
"And the Dialogue Experts’ knowledge about products is vast and
available instantly, at any time"
"Shoppers benefit from an enhanced user experience, better service
and instant response time"
"E-commerce websites benefit from higher sales, lower costs and
invaluable market intelligence gained from the Dialogue Mining of
shoppers’ conversations"
Self-Description of Soliloquy (2)
110
Clients
• The example dialogs with the Notebook Expert that are presented
on the following slides were conducted at http://www.buy.com in
December, 2000
• At that time, the Notebook Expert was also deployed at
http://www.cnet.com
• At both sites, the Notebook Expert has apparently since been taken
out of the site
• Similarly, other earlier clients of Soliloquy (Acer, Hewlett-Packard,
and Hardware Street) have apparently discontinued use of the
Notebook Expert
111
Using Lifelike Characters / The Notebook Expert
112
Examples of Possible Inputs
111
According to the web site, the Notebook Expert can deal with user
inputs of the following types, among others:
Statements about the
planned use of the system
Specification of preferences
• I want a cheap notebook
• I travel a lot
• I plan to spend around
• I need a laptop for school
$2000
• Show me a notebook with a
big screen
• Show me a really fast laptop
computer
• I need a laptop for
•
•
•
•
•
accounting
I do desktop publishing
I trade in stocks, shares,
options and futures
I play games on the
computer
I need a laptop for graphics
Show me one for
spreadsheets
Questions about concepts
• What is USB?
• What is a sound card?
Dialog With the Notebook Expert (1)
When U arrives at buy.com’s main notebook page, she can sees the
link (on the right) to the page of the Notebook Expert
112
113
Using Lifelike Characters / The Notebook Expert
113
Dialog With the Notebook Expert (2)
114
Here, S starts by asking about the features that U would like
In some cases, S starts with a question about how U intends to use the
notebook (see the examples above)
Dialog With the Notebook Expert (3)
As we can see in the Quick Summary on the right, the vague
specification "a lot of memory" (see the previous slide) was translated
into the internal specification "> 64 Megabyte"
114
115
Using Lifelike Characters / The Notebook Expert
115
Dialog With the Notebook Expert (4)
116
The value "$3000" in the vague specification "I can spend about $3000"
(previous slide) was interpreted not as an upper limit but as a desired
price
At any time, U can click on one of the hyperlinks in the window listing
the notebooks to go to a web page about that specific notebook
Dialog With the Notebook Expert (5)
116
Here we can see how S interpreted the vague specification "I don’t need
more than 600 mhz"
117
Using Lifelike Characters / The Notebook Expert
117
Dialog With the Notebook Expert (6)
118
Here is an example of an attempt by S to answer a question by U about
a particular concept (here: a particular manufacturer)
Dialog With the Notebook Expert (7)
Here is another example of S’s natural language understanding
capabilities
118
119
Using Lifelike Characters / The Notebook Expert
119
Dialog With the Notebook Expert (8)
120
S is actually capable of processing some statements of this general sort
E.g.,. "I go to conferences a lot"
It would be unreasonable to expect S to be able to understand every
possible statement of this type, since this capability would require a
great deal of world knowledge
Still, an interpretation problem should obviously not have such drastic
consequences for the course of the dialog
Artificial Life
Bot on Artificial Life Main Page
120
121
Using Lifelike Characters / Artificial Life
122
Advertised Functions of Bots
121
From a product brief available from http://www.artificial-life.com.
What Can a Bot Do for You?
• Create individual customer profiles, which includes information
customers are seeking
• Provide detailed information to your customer that leads to sales
• Facilitate customer awareness regarding products and services
• Offer a patient and insightful presence which improves customer
loyalty
• Enhance the customer’s experience through productive, natural
language conversations with your bot
Example Dialog (1)
122
• The following slides show a dialog conducted with Luci
• The bot’s output (small white letters) appears before the user types
the next input into the white field above it
• The gestures and varying facial expressions of the bot are not
captured in these screen shots, because they appear only briefly
before she returns to her standard pose and expression
123
123
Using Lifelike Characters / Artificial Life
124
Example Dialog (2)
Example Dialog (3)
124
125
125
Using Lifelike Characters / Artificial Life
126
Example Dialog (4)
Example Dialog (5)
126
127
127
Using Lifelike Characters / Artificial Life
128
Example Dialog (6)
Example Dialog (7)
Here, instead of typing a question, the user clicks on the link "Einstein
ALife"
128
129
Using Lifelike Characters / Artificial Life
130
Example Dialog (8)
129
When the new page appears, Luci introduces its main character
The PPP Persona
More information on PPP, including the publications cited below, is available
from http://www.dfki.de/imedia/ppp/.
Overview of PPP
130
History
• In a series of research projects at DFKI since 1989, techniques for
automatic interactive presentations have been developed
• In the project PPP ("Personalized Plan-Based Presenter"), one of
the first "personas" was introduced in 1994
• This line of research is now being pursued in the project Miau (see
below)
Relationship to other personas
• The PPP persona employs an especially large repertoire of
behaviors
• Its behaviors are planned for large presentation units
• Since the focus is on presentation techniques, in most cases no
language input of the user is processed
131
132
PPP: Example (1)
André, E., Rist, T., & Müller, J. (1998). Guiding the user through dynamically
generated hypermedia presentations with a life-like character. In J. Marks
(Ed.), IUI98: International Conference on Intelligent User Interfaces
(pp. 21−28).New York: ACM.
131
Using Lifelike Characters / The PPP Persona
PPP: Example (2)
•
•
•
•
•
132
U wants to spend holidays in Finland and is looking for a lakeside
cottage
S retrieves matching offers from the WWW, selects one of them, and
presents it to U
So that U can ask for more information, several items in the text are
mouse−sensitive
• Clicking on one of these items will lead to the insertion of a
subscenario
For instance, if the user clicks on the fishing item while the first
cottage is presented, S will interrupt the current presentation and run
a script with fishing possibilities
• After that, S will continue with the main script and describe the
next offer
Note that following a navigation link does not cause paging, as it
does with most conventional web presentations
• Rather, a new presentation script for the agent along with the
required textual and pictorial material is transferred to the
client−sidepresentation runtime engine
133
133
Using Lifelike Characters / The PPP Persona
134
Self-Behaviors (1)
Self-Behaviors (2)
134
• The persona’s primary purpose is to execute presentation acts
• But its behavior is not only determined by the directives (i.e.,
presentation tasks) specified in the script
• In addition, self−behaviors are indispensable in order to increase the
Persona’s vividness and believability
• Though it is certainly possible to include appropriate instructions
directly in the presentation script, the approach here is to have them
determined automatically
135
135
Using Lifelike Characters / The PPP Persona
136
Automatic Generation of Presentations
• Since the available cottages and also their features may change at
any time, it doesn’t make sense to rely on predesigned
presentations
• The generation process is therfore automated; it comprises the
following tasks:
1. the design of a multimedia discourse structure reflecting how the
single parts of a presentation are related to each other
2. the decomposition of the presentation into self−contained
presentation units
3. the design of a navigation graph
4. the design of presentation scripts for each presentation unit
See, among others, André, E., Rist, T., & Müller, J. (1999). Employing AI
methods to control the behavior of animated interface agents. Applied
Artificial Intelligence.
Experimental Evaluation
136
Variants that were compared
1. Presentations with the persona, as was explained above
2. The same presentations without the visible persona, only with
speech output and pointing with a cursor
Main results
3. The presence of a persona invoked positive affective reactions in
the users
4. The objectively measurable uptake of information was neither better
nor worse with the persona
• One possible explanation is that, in this study, the behavior of the
persona conveyed little information beyond mere pointing and
speaking
Conclusion
• An animated human-like figure can improve subjective acceptance −
without distracting users from the content of the presentation
137
Using Lifelike Characters / The MIAU Agents
138
The MIAU Agents
Overview of MIAU
MIAU-Homepage: http://www.dfki.de/imedia/miau/. Publication: Elisabeth
André, T. R., (2000). Presenting through performing: On the use of multiple
lifelike characters in knowledge-based presentation systems. In H.
Lieberman (Ed.), IUI 2000: International Conference on Intelligent User
Interfaces (p. 1). New York: ACM.
http://lieber.www.media.mit.edu/people/lieber/IUI/
137
Basic ideas
• The main concept in MIAU is that of a team of several agents that
converse about a product that U is interested in
Noninteractive MIAU
• In the initial version, U does not communicate with the agents
• Instead, U watches the agents converse
• But U can set the parameters that determine the interaction
Interactive MIAU
• In a version currently under development (shown first in this tutorial,
with a video) U can direct utterances to the agents
Advantages claimed for the use of multiple agents
• Different points of view and types of information can often be
presented especially effectively by different agents
• Repetitions of claims (e.g., about the advantages of a product) are
less boring to U when they are generated by different agents
Interactive MIAU: Examples (1)
Robby (left) is helping out by answering the more technical questions;
he has just reported the (high) gas consumption of the car
138
139
139
Using Lifelike Characters / The MIAU Agents
140
Interactive MIAU: Examples (2)
Because of his personality parameters, Robby expresses agreement
with U’s negative assessment
Interactive MIAU: Examples (3)
Merlin, who has the personality parameters of a normal salesman,
expresses his dismay and disbelief through gestures and facial
expressions
140
141
141
Using Lifelike Characters / The MIAU Agents
142
Setting the Presentation Parameters
In this screen, U determines which agents are to appear and what
general sort of behavior they are supposed to exhibit
Noninteractive MIAU: Examples (1)
142
The potential customer, here played by the agent Genie (right) has just
expressed some concerns about the car in question; the salesman
Merlin (right) attempts to downplay these concerns
143
143
Using Lifelike Characters / The MIAU Agents
144
Noninteractive MIAU: Examples (2)
The next two slides show an example of a dialog that was conducted
with the following agents and parameters:
Agent
Role
Personality Factors
Interests
Robby
Seller
Extraverted, agreeable
Sportiness
Peedy
Buyer
Introverted, disagreeable Environment
Merlin
Buyer
Extraverted, agreeable
Safety
Noninteractive MIAU: Examples (3)
144
Agent
Utterance
Commentary
Robby:
Hello, I’m Robby. What
can I do for you?
We are interested in this
car.
This is a very sporty car.
It can drive 100 miles per
hour.
Initiates the conversation, because
of extraversion
Answers the question, because of
extraversion
Emphasizes a value dimension that
is important to U; mentions an
attribute that has positive
implications with respect to this
dimension
Asks questions because of
extraversion; wants to know about
airbags because this attribute
concerns the dimension of safety,
which is important to him
Merlin:
Robby:
Merlin:
Does it have airbags?
Robby:
Merlin:
Sure.
Excellent.
Positive evaluation, because the
agent is "agreeable"; strong
formulation because of extraversion
145
Using Lifelike Characters / The MIAU Agents
146
Noninteractive MIAU: Examples (4)
145
Agent
...
Peedy:
Robby:
Peedy:
Robby:
Utterance
Commentary
How much gas does it
consume?
It consumes 8 liters per
100 km.
Isn’t that bad for the
environment?
Bad for the
environment? It has a
catalytic converter. It is
made of recyclable
material.
Gas consumption has implications
with respect to the environment
Negative comment because the
agent is "disagreeable"
REA
Image from http://justine.www.media.mit.edu/people/justine/.
REA: An Embodied Information Presenter
146
147
Using Lifelike Characters / REA
Example Dialog (1)
147
Cassell, J. (2001). Embodied conversational agents: Representation and
intelligence in user interface. AI Magazine, 22.
http://justine.www.media.mit.edu/people/justine/publications.html
148
• The following records an actual interaction between a user and Rea
• Mike approaches the projection screen. Rea is currently turned side
on and is gazing idly about
• As Mike moves within range of the two cameras mounted above the
screen, Rea turns to face him
REA: Hello. How can I help you?
MIKE: I’m looking to buy a place near MIT.
Rea nods, indicating that she is following.
REA: I have a house to show you.
A picture of a house appears on-screen behind Rea who blinks
and looks at the house and then at Mike.
REA: It’s in Somerville.
MIKE: Tell me about it.
Rea looks up and away while she plans what to say
Example Dialog (2)
148
REA: It’s big.
Rea makes an expansive gesture with her hands.
Mike brings his hands up as if he is about to speak and so Rea
does not continue; instead waiting for him to speak.
MIKE: Tell me more about it
REA: Sure thing. It has a nice garden.
Rea sketches a curved gesture with her hands indicating that the
garden extends along two sides of the house
MIKE: How far is it?
REA: It is five minutes to the Porter Square T station.
Rea makes it clear that it is five minutes on foot from the T station
by making a walking gesture with her fingers
MIKE: How big is the house?
149
Using Lifelike Characters / REA
150
Example Dialog (3)
149
REA: It has four bedrooms, three bathrooms. . .
Mike interrupts Rea who stops speaking immediately.
MIKE: Wait. Tell me, where is the master bedroom?
REA: I’m sorry, I didn’t catch that. What did you ask me?
MIKE: Where is the master bedroom?
REA: It’s upstairs.
Rea points up
MIKE: Where is the master bathroom?
REA: It’s next to the bedroom.
Rea brings her hands together to indicate the relationship
between the bedroom and the bathroom.
And the house tour continues ...
Cassell, J., Bickmore, T., Vilhjálmsson, H., & Yan, H. (2000). More than just
a pretty face: Affordances of embodiment. In H. Lieberman (Ed.), IUI 2000:
International Conference on Intelligent User Interfaces (pp. 52−59). New
York: ACM. http://lieber.www.media.mit.edu/people/lieber/IUI/
Functions of Nonverbal Behaviors (1)
150
Conversational functions and behavioral realizations (1)
Communicative Functions
Communicative Behavior
Initiation and termination
Reacting
Short glance
Inviting contact
Sustained glance, Smile
Distance salutation
Looking, Head toss/nod, Raise
eyebrows, Wave, Smile
Close salutation
Looking, Head nod, Embrace or
handshake, Smile
Break away
Glance around
Farewell
Looking, Head nod, Wave
151
151
Using Lifelike Characters / REA
152
Functions of Nonverbal Behaviors (2)
Conversational functions and behavioral realizations (2)
Communicative Functions
Communicative Behavior
Turn-Taking
Give turn
Looking, Raise eyebrows (followed by
silence)
Wanting turn
Raise hands into gesture space
Take Turn
Glance away, Start talking
Feedback
Request feedback
Looking, Raise eyebrows
Give feedback
Looking, Head nod
Comparison
Comparison of Approaches (1)
Soliloquy
ALife-WebGuide
PPP
MIAU
152
REA
1. To what extent does U have to install additional software on her own computer?
No
No
No
Microsoft Agents;
interactive version
not yet
web-capable
System exists only
as prototype
2. What sort of language or speech generation is S capable of?
Filling in of
templates
Canned texts
Reading aloud of
pre-stored texts
Reading aloud of
texts generated
with scripts
Generation of
speech coordinated
with nonverbal
behaviors
3. What sorts of input can the persona understand, aside from typed text?
None
None
(No free language
input); selection
from presented
questions and
evaluations
Specification of
dialog parameters
via user interface
Speech and
nonverbal
behaviors
153
153
Soliloquy
Using Lifelike Characters / Comparison
154
Comparison of Approaches (2)
ALife-WebGuide
PPP
MIAU
REA
Facial expressions
and gestures as
supported by
Microsoft agents
Facial expressions
and gestures
4. What means of emotional expression does S have?
Only verbal
expressions
Sometimes
appropriate facial
expressions
Dynamically
generated facial
expressions and
gestures
5. What other nonverbal means of expression does S have?
Dynamically
generated tables
Glances toward
relevant parts of
screen
Pointing
coordinated with
speaking
Occasional
Gestures with
meaningful
considerable
gestures (e.g.,
semantic content
looking into a book)
6. To what extent is the behavior of S coordinated with the other aspects of U’s
interaction with S?
U can pursue links
to other pages
Often helpful
comments that
complement web
pages
Tight linking
between persona’s
behavior and other
elements of the
presentation
(There are no other S’s behavior is
aspects of the
coordinated with
interaction)
display of pictures
Comparison of Approaches (3)
Soliloquy
ALife-WebGuide
PPP
MIAU
154
REA
7. To what extent does S help U to concentrate her attention on the pursuit of her goal?
Excessive
structuring of the
dialog
OK, when
understanding is
successful
Strong focusing,
despite U’s
freedom of action
Complete focus on
the product
discussion, no
other options for U
Complete focus on
the product
discussion, no
other options for U
8. To what extent is U made to feel that she is being recognized and treated as an
individual?
Longer-term but
simple adaptation
Only short-term
adaptation to U
Some longer-term
adaptation, but not
always visible to U
Agents’ utterances
are responsive to
U’s
Fine-grained
responses to U’s
behavior