Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
User Modeling Meets Usability Goals
Anthony Jameson
DFKI, German Research Center for Artificial Intelligence and
International University in Germany
3 Title Page 4
Contents
3
Introduction 5 A Haunting Question 5 This Is Not All New ... 6 What Are the Messages of This Talk? 7
Goals and Typical Threats 8 Controllability 8 Comprehensibility 9 Unobtrusiveness 10 System Competence 11 Privacy 12 Breadth of Experience 13
Controllability vs. Obtrusiveness 14 Intelligent Office System 14 Early Version of Confirmation Prompt 15 Prompt on the Touch Screen 16 Control Panel 17 Causes and Strategies 18 Expanding the Design Space 19
Breadth of Experience vs. System Competence 20 A Decision−Theoretic Shopping Guide 20 Direction to Walk In 23 Overview Map 24 Photo of Upcoming Store 25 Study in Shopping Mall: Method 26 Objective Results 27 Subjective Results 28 Breadth of Experience 29 Causes and Strategies 30 Expanding the Design Space 31
Control and Comprehension vs. Obtrusiveness 32 Control and Comprehension 32 Causes and Strategies 33 Explanations: Implementation 34 Explanations: When Presented 35 Explanations: Results 36 Expanding the Design Space 37
Comprehensibility vs. Obtrusiveness 38 An Adaptive Hotlist for Conference Events 38 Overview of Studies 39 Causes and Strategies 40 Comprehensibility of the Hotlist 41 Impact of Explanations 42 Expanding the Design Space 43
Controllability vs. System Competence 44 Causes and Strategies 44 Some Results (Study 1) 45 Advantages of Two Updating Styles 46 Improved Interface 48 Some Results (Study 2) 49 Expanding the Design Space 50
Concluding Remarks 51 The Messages Again 51 What Does the Title Mean? 52
5 Introduction 6
Introduction A Haunting Question
5
When my fancy novel techniques
finally work well enough to be
used in real systems . . .
will anyone want to use
these systems?
This Is Not All New ...
6
Usability threats and principles
• Ben Shneiderman, since mid−1990s
• Pattie Maes and coworkers, late 1990s
• Eric Horvitz, 1999
• Kristina Höök, 2000
• ...
Evaluation of user−adaptive systems
• David Chin
• Stephan Weibelzahl
• Alexandros Paramythis
• Judith Masthoff
• ...
7 Introduction 8
What Are the Messages of This Talk?
7
1. User−adaptivity is
fundamentally a great way to increase the usability of interactive systems
2. Just apply general guidelines like "Put the user in control"
3. User modeling is an alternative paradigm to mainstream human−computer interaction paradigms
The wrong messages The real messages 1. User−adaptivity
requires careful analysis of typical usability threats
2. Because of tradeoffs, no single solution is right for all of the users all of the time
3. By expanding the design space, you can find ways to satisfy more of the users more of the time
Goals and Typical Threats Controllability
8 A d
iscu
ssio
n of
thes
e go
als
and
thre
ats
will
be
foun
d in
Sec
tion
4 of
: Ja
mes
on, A
. (20
03).
Ada
ptiv
e in
terfa
ces
and
agen
ts. I
n J.
Jac
ko &
A. S
ears
(E
ds.),
Hum
an−c
ompu
ter i
nter
actio
n ha
ndbo
ok (p
p. 3
05−3
30).
Mah
wah
, NJ:
E
rlbau
m. A
revi
sed
vers
ion
is b
eing
pre
pare
d fo
r the
2nd
edi
tion,
sch
edul
ed
for 2
006.
The user may not have enough control over the system
9 Goals and Typical Threats 10
Comprehensibility
9
The user may not understand adequately how the system works −or be able to predict what it will do
Unobtrusiveness
10
!
??
The system may distract the user with too many (or poorly timed) messages and requests for input
11 Goals and Typical Threats 12
System Competence
11
The system may perform actions that are so poorly adapted to actual facts about the user that the user is distracted and/or impeded
Privacy
12 Priv
acy
is n
ot d
iscu
ssed
in th
is ta
lk, b
ecau
se it
is th
e su
bjec
t of t
he in
vite
d ta
lk a
t thi
s co
nfer
ence
by
Lorr
ie F
aith
Cra
nor
The system may create situations in which information that the user would prefer to keep private are made available to others
13 Goals and Typical Threats 14
Breadth of Experience
13
The system may restrict the user’s attention excessively
Controllability vs. Obtrusiveness Intelligent Office System
14 Che
vers
t, K
., B
yun,
H. E
., Fi
tton,
D.,
Sas
, C.,
Kra
y, C
., &
Vill
ar, N
. (20
05).
Exp
lorin
g is
sues
of u
ser m
odel
tran
spar
ency
and
pro
activ
e be
havi
our i
n an
of
fice
envi
ronm
ent c
ontro
l sys
tem
. Use
r Mod
elin
g an
d U
ser−
Ada
pted
In
tera
ctio
n. In
pre
ss.
(Cheverst et al., UMUAI special issue on User Modeling in Ubiquitous Computing)
15 Controllability vs. Obtrusiveness 16
Early Version of Confirmation Prompt
15
On user’s main workstation window:
Prompt on the Touch Screen
16 The
wor
d "O
FF"
chan
ges
colo
r rep
eate
dly
whi
le th
e pr
ompt
is b
eing
sho
wn
17 Controllability vs. Obtrusiveness 18
Control Panel
17
Causes and Strategies
18
!
??
���������
���������
Allow user todo some of the work
���������
���������
Craft imageof systemcarefully
!Taking over bysystem of workof user
!Anthropo−morphicappearance
RemedialMeasures
UsabilityThreatsMeasures
PreventiveCausesTypical
Submit systemactions to userfor approval
Allow userto set systemparameters
Shift controlto systemgradually
Design messa−ges carefully(form, timing)
19 Controllability vs. Obtrusiveness 20
Expanding the Design Space
19
!
??
!
??
!
??
!
??
Preemptoryquery
actionAutonomous
Nonpreemptoryrecommendationto press button
Autonomousaction inspecifiedsituations
Announcementof action withinN seconds unlesscanceled
Breadth of Experience vs. System Competence A Decision−Theoretic Shopping Guide (1)
20 The
shop
ping
gui
de a
nd u
ser s
tudy
sho
wn
in th
e sl
ides
in th
is a
nd th
e fo
llow
ing
sect
ion
are
pres
ente
d in
: Boh
nenb
erge
r, T.
, Jac
obs,
O.,
Jam
eson
, A
., &
Asl
an, I
. (20
05).
Dec
isio
n−th
eore
tic p
lann
ing
mee
ts u
ser r
equi
rem
ents
: E
nhan
cem
ents
and
stu
dies
of a
n in
telli
gent
sho
ppin
g gu
ide.
In H
. Gel
lers
en,
R. W
ant,
& A
. Sch
mid
t (E
ds.),
Per
vasi
ve c
ompu
ting:
Thi
rd in
tern
atio
nal
conf
eren
ce (p
p. 2
79−2
96).
Ber
lin: S
prin
ger.
21 Breadth of Experience vs. System Competence 22
A Decision−Theoretic Shopping Guide (2)
21
A Decision−Theoretic Shopping Guide (3)
22
The decision−theoretic shopping guide • The shopper specifies at the beginning her
interests in particular (types of) products • "A loaf of pumpkin seed bread" • "A novel for my teen−aged daughter" • ...
• The system computes a policy: • At each point in time, it directs the shopper to a
promising store, taking into account: 1. the current location 2. the products found so far 3. the amount of time remaining
23 Breadth of Experience vs. System Competence 24
Direction to Walk In
23
Overview Map
24
25 Breadth of Experience vs. System Competence 26
Photo of Upcoming Store
25
Study in Shopping Mall: Method
26
• The localization infrastructure was simulated by the experimenter (Wizard of Oz)
• 21 subjects from different social groups
• Each shopped for 20 minutes with 25 Euros after specifying what they wanted to buy in six categories:
• Some bread, a book, a gift item, some fruit, a magazine, some stationery
27 Breadth of Experience vs. System Competence 28
Objective Results
27
(b) Time to finish despitenot having bought all 6 items
(a) Time needed to buy all 6 items
< 14 18 to 2016 to 1814 to 1615 to 20< 10 10 to 15Time (minutes)Time (minutes)
12
0
8
4
Num
ber o
f sub
ject
s 12
0
8
4
• All 21 subjects got back to the exit on time
Subjective Results
28
15
0
10
5
20
15
0
10
5
20
No Maybe YesProbably notYesA littleNo
NoneA littleNot at all A lot Some ManyA few
15
0
10
5
20
15
0
10
5
20
Num
ber o
f sub
ject
sN
umbe
r of s
ubje
cts
(c) Feeling restricted (d) Difficulties
(a) Enjoyment (b) Willingness to use
29 Breadth of Experience vs. System Competence 30
Breadth of Experience
29
Critique • "Shoppers don’t like to be led around on a fixed
route • They want to explore and buy spontaneously and
have fun while doing so"
Response • Not all shoppers are the same all of the time • Our subjects expressed interest in using the
system when ... • ... they are unfamiliar with the shopping mall • ... they want to buy a particular set of products • ... their time is limited
Causes and Strategies
30
������������
���������
Allow user todo some of the work
������������
���
Acquire a lot of relevant information
!Taking over bysystem of workof user
!Incompletenessof system’sinformation
Intentionallyintroducediversity
actionsthe system’sExplain
MeasuresPreventive
CausesTypical Remedial
MeasuresUsabilityThreats
31 Breadth of Experience vs. System Competence 32
Expanding the Design Space
31
Control and Comprehension vs. Obtrusiveness Control and Comprehension
32 This
sec
tion
was
not
incl
uded
in th
e pr
esen
tatio
n at
UM
200
5, b
ecau
se o
f th
e tim
e lim
itatio
n
• Why do users sometimes want more control and understanding? • So that they can override the system’s
recommendations They have information that the system lacks They see that the system’s model is too limited
• What do they need? • Robust response by the system when they
deviate from a recommendation ⇒ Given by the basic algorithm
• Ability to second−guess the system in an informed way ⇒ Requires explanations by the system
33 Control and Comprehension vs. Obtrusiveness 34
Causes and Strategies
33
Shift controlto systemgradually
actionsthe system’sExplain
Allow inspect−ion of theuser model
Submit systemactions to userfor approval
Allow userto set systemparameters
���
���������
Use simplemodelingtechniques
������������
� � �
Acquire a lot of relevant information
������������
���������
Allow user todo some of the work
RemedialMeasures
UsabilityThreatsMeasures
PreventiveCausesTypical
! processingComplex
!Taking over bysystem of workof user
!Incompletenessof system’sinformation
Explanations: Implementation
34
35 Control and Comprehension vs. Obtrusiveness 36
Explanations: When Presented
35
Difference between best and second−best options:
Small Medium Large
Explanations: Results
36
Results • Five subjects used the system with explanations • They generally approved of the basic idea • But most said that they had too little time to look at
the explanations and preferred to follow the recommendations blindly
Prediction • With more experience, each user would learn in
what situations it is worthwhile to check th explanation • E.g., when they are tempted to second−guess
the system
37 Control and Comprehension vs. Obtrusiveness 38
Expanding the Design Space
37
!
??
!
??
!
??
!
??
Pre− or postshoppingcritique (cf. SPECTER)
Comprehensibility vs. Obtrusiveness An Adaptive Hotlist for Conference Events
38 This
sys
tem
was
dem
onst
rate
d liv
e du
ring
the
pres
enta
tion
at U
M 2
005.
It is
us
ually
acc
essi
ble
via
http
://df
ki.d
e/um
2001
. Thi
s sc
reen
sho
t sho
ws
an
earli
er v
ersi
on, w
hich
was
use
d fo
r the
con
fern
ce it
self
and
for S
tudy
1.
39 Comprehensibility vs. Obtrusiveness 40
Overview of Studies
39 Stu
dy 1
is re
porte
d in
: Jam
eson
, A.,
& S
chw
arzk
opf,
E. (
2002
). P
ros
and
cons
of c
ontro
llabi
lity:
An
empi
rical
stu
dy. I
n P
. De
Bra
, P. B
rusi
lovs
ky, &
R.
Con
ejo
(Eds
.), A
dapt
ive
hype
rmed
ia a
nd a
dapt
ive
web
−bas
ed s
yste
ms:
P
roce
edin
gs o
f AH
200
2 (p
p. 1
93−2
02).
Ber
lin: S
prin
ger.
http
://df
ki.d
e/~j
ames
on/ A
pub
licat
ion
that
incl
udes
a re
port
on S
tudy
2 is
cu
rren
tly in
pre
para
tion.
• Experiment with original version (see previous slide) • 18 student subjects
Made to act like UM researchers (How? ⇒ Discussion)
Comparison between controlled and automatic updating
• Experiment with improved (current) version • Same as above, but:
28 student subjects 12 without the ++s and −−s
Causes and Strategies
40
MeasuresPreventive
CausesTypical Usability
ThreatsRemedialMeasures
actionsthe system’sExplain
Submit systemactions to userfor approval
Shift controlto systemgradually
! processingComplex
!Incompletenessof system’sinformation
!
??
������������
���������
Acquire a lot of relevant information
���������
���������
Use simplemodelingtechniques
Allow inspect−ion of theuser model
!Implicit inform−
tion by systemation acquisi−
���������
���������
about informa−Inform user
tion acquisition
41 Comprehensibility vs. Obtrusiveness 42
Comprehensibility of the Hotlist
41
• Theory: The explanations can help the user to understand ... • Why this particular recommendation was made • What the system’s basic procedure for making
recommendations is • How accurate the system’s user model is at the
present time • The user should then be better able to predict
• Whether this particular event will turn out to be interesting to the user
• What sorts of recommendations the system will make in the future
• How valuable these recommendations will be
Impact of Explanations
42
• Those with explanations did a bit better (p < .05) on a "comprehension test":
"Does the system take into account ...
... 1. what talks you have added to the hotlist? [correct: ’Yes’]
... 2. what pages you have looked at? [’Yes’]
... 3. how long you looked at each page? [’No’]"
• Most found them "somewhat useful" or "useful to a small extent"
43 Comprehensibility vs. Obtrusiveness 44
Expanding the Design Space
43
!
??
!
??
!
??
!
??
[Title, authors]
[Title, authors]Machine Learning,Empirical Studies
[Title, authors]Machine Learning (−−−),Empirical Studies (+)
[Full explanation.]
How Does the Hotlist Work?
Controllability vs. System Competence Causes and Strategies
44
!
??
UsabilityThreats
RemedialMeasuresMeasures
PreventiveCausesTypical
Submit systemactions to userfor approval
Allow userto set systemparameters
Shift controlto systemgradually
������������
���������
Allow user todo some of the work
!Taking over bysystem of workof user
45 Controllability vs. System Competence 46
Some Results (Study 1)
45
Probablywoulduse
Probablywoulduse
Quitewillingto use
Probablywould
not use
Wouldnotcare
Probablywould
not use
Woulddefinitelynot use
Veryeagerto use
Wouldnotuse
Wouldnotcare
Probablywoulduse
R
F
With controlled updating?
C
I J
D E
K
M
LP Q A
G H B
N
O
With
aut
omat
ic u
pdat
ing?
Woulddefinitelynot use
Woulddefinitelynot use
Wouldnotuse
Wouldnotcare
Probablywoulduse
Probablywould
not use
Veryeagerto use
Quitewillingto use
Wouldnotuse
Probablywould
not use
Wouldnotcare
Quitewillingto use
Probablywoulduse
Veryeagerto use
Woulddefinitelynot use
Veryeagerto use
Wouldnotuse
Quitewillingto use
Probablywould
not use
Probablywould
not use
Advantages of Two Updating Styles (1)
46
Controlled updating: 1. The user’s feeling of control over the interaction
with the system is enhanced
2. The user can follow up on more than one recommendation in a given set
3. System response times can be faster because of less frequent updating
4. The user can restrict updates to situations in which the system’s model of her interests is assumed to have useful accuracy
5. A smaller amount of irrelevant text appears in the hotlist.
47 Controllability vs. System Competence 48
Advantages of Two Updating Styles (2)
47
Automatic updating:
1. The user cannot overlook the availability of the recommendation feature
2. The user is regularly reminded that new recommendations are available
3. The user is spared the effort of clicking on a button to obtain new recommendations
4. The recommendations displayed always reflect the system’s most complete model of the user’s interests
Improved Interface
48 This
inte
rface
ach
ieve
s th
e se
cond
adv
anta
ge o
f con
trolle
d up
datin
g (s
ee
the
earli
er s
lide)
whi
le s
till a
llow
ing
auto
mat
ic u
pdat
ing.
(With
aut
omat
ic
upda
ting,
the
butto
n "U
pdat
e re
com
men
datio
ns"
is n
ot a
vaila
ble.
)
49 Controllability vs. System Competence 50
Some Results (Study 2)
49
• Some drawbacks of automatic updating were eliminated through the interface improvements
• Preferences generally shifted toward automatic updating
• But there were still large differences in preferences concerning almost all aspects of the interaction
Expanding the Design Space
50
!
??
!
??
!
??
!
??
[Automatic updating]
Update recommendations
Execute changes
[Automatic updating]
Update recommendations
Execute changes
51 Concluding Remarks 52
Concluding Remarks The Messages Again
51
1. User−adaptivity requires careful analysis of typical usability threats
2. Because of tradeoffs, no single solution is right for all of the users all of the time
3. By expanding the design space, you can find ways to satisfy more of the users more of the time
What Does the Title Mean?
52
Usability GoalsUser Modeling Meets
Gets to Know Confronts
Achieves