3. PEOPLE DIFFER IN WHAT
THEY LIKE
recommender system: computing .......
likes blue
likes purple
4. MORE DIFFERENCES...
Recommendations are tailored...
...but the interaction is the same for every user
People also differ in how they make
decisions!
They may need different ways of interacting with
the system
5. TAILORED INTERFACES
We take a closer look at these
differences...
On what characteristics do users differ?
Do they use different decision strategies?
...and tailor the system to these
differences
Which interaction methods support these
strategies?
6. DECISION STRATEGIES
A decision strategy
is a procedure for making decisions
Weigh the attribute values
Pick the very first item you see
Different interaction methods may
support these strategies to a different
extent
7. PERSONAL DIFFERENCES
The decision strategy selected by the
user depends on her personal
characteristics
Bettman, Luce & Payne, 1998
Typical characteristics:
Experts vs. Novices (Alba & Hutchinson, 1987; Coupey et al.,
1998)
Distrusting vs. Trusting (Vries, 2004; Wang & Benbasat,
2007)
8. IN SHORT...
Different user
Different decision
Different interaction
16. FIVE INTERACTION METHODS
TopN: the 10 most popular measures
Baseline condition; not personalized, virtually no
interaction
Get more recommendations by classifying measures
17. FIVE INTERACTION METHODS
Sort: sort the measures by any attribute
Implements the lexicographic strategy:
‣ Select the most important attribute
‣ Choose the item with the highest value on this attribute
Users can (re-)sort by clicking on table headers
18. FIVE INTERACTION METHODS
Explicit: a typical MAUT recommender
Implements the weighted adding strategy
‣ Normalize the attribute values: vij
‣ Assign weights to the attributes: wj
‣ Multiply and sum to get a utility: Ui = ∑vij* wj
‣ Choose item with highest utility
19. FIVE INTERACTION METHODS
Implicit: system decides on the weights
User behavior is analyzed to update the weights
Update rules based on previous versions of the
system
Weights are not shown
21. PARTICIPANTS
147 participants
(158 at first, 11 removed due to very short
interaction time)
Recruited by an external company
79 male, 68 female
Average age: 40 (sd: 15.9)
29 students, 93 working, 25 retired
23 high school, 24 intermediate degree, 53 college,
47 grad
22. PERSONAL CHARACTERISTICS
Domain knowledge: Experts vs. Novices
‣ 7 items, e.g. “I understand the difference between energy
saving measures”
Trusting propensity: Distrusting vs.
Trusting
‣ 6 items, e.g. “In general, most folks keep their promises”
Persistence: Satisficers vs. Maximizers
‣ 4 items, e.g. “I am willing to examine the product
attributes very carefully in order to make sure that the
23. USER EXPERIENCE
Control: Does it support my strategy?
‣ 7 items, e.g. “I had full control over the system”
Understandability: Is it confusing?
‣ 8 items, e.g. “I understand the system”
Trust in the system: Is it fair to me?
‣ 4 items, “The system is not biased”
24. USER EXPERIENCE
QUIS: Is the user interface usable?
‣ 5 items, summed 9-point scale
Perceived system effectiveness: Is it
useful?
‣ 5 items, “I make better choices with this system”
Choice satisfaction: Do I like what I
chose?
‣ 4 items, “I think I chose the best measures”
27. DOMAIN KNOWLEDGE
Novices may like TopN,
Sort and Implicit
because they lack attribute
knowledge
Implicit may be more confusing
Experts may like Explicit
and Hybrid
because they can leverage their
attribute knowledge, and because
28. DOMAIN KNOWLEDGE
Novices like 2!
the TopN *!
system
1!
TopN!
Control!
Sort!
0!
-2! -1! 0! 1! 2! Explicit!
Implicit!
Hybrid!
-1!
-2!
Domain Knowledge!
29. DOMAIN KNOWLEDGE
Novices like 2!
the TopN *!
system
1!
TopN!
They perceive
Control!
Sort!
more control in -2! -1!
0!
0! 1! 2! Explicit!
this system than Implicit!
experts -1!
Hybrid!
-2!
Domain Knowledge!
30. DOMAIN KNOWLEDGE
Novices like
2!
***!
Perceived system effectiveness!
the TopN
system
1! 1! **!
TopN!
They perceive Sort!
more control in -2! -1!
0!
0! 1! 2! Explicit!
this system than Implicit!
experts
Hybrid!
-1!
They find it far
more effective
-2!
Domain Knowledge!
31. DOMAIN KNOWLEDGE
Novices like
2!
***!
Perceived system effectiveness!
the TopN
system
1! 1! **!
TopN!
They perceive Sort!
more control in -2! -1!
0!
0! 1! 2! Explicit!
this system than Implicit!
experts
Hybrid!
-1!
They find it far
more effective
-2!
Domain Knowledge!
33. DOMAIN KNOWLEDGE
Experts like the 2!
Hybrid system
Understandability!
1!
*!
TopN!
Sort!
0!
-2! -1! 0! 1! 2! Explicit!
Implicit!
Hybrid!
-1!
-2!
Domain Knowledge!
34. DOMAIN KNOWLEDGE
Experts like the 2!
Hybrid system
They understand it Understandability!
1!
*!
TopN!
Sort!
0!
-2! -1! 0! 1! 2! Explicit!
Implicit!
Hybrid!
-1!
-2!
Domain Knowledge!
35. DOMAIN KNOWLEDGE
Experts like the 45!
Hybrid system
40!
User interface satisfaction!
35!
They understand it
30!
***!
They are more 25!
TopN!
Sort!
satisfied with its UI Explicit!
20!
**!
Implicit!
15! Hybrid!
10!
5!
-2! -1! 0! 1! 2!
Domain Knowledge!
36. DOMAIN KNOWLEDGE
Experts like the ***!
2!
Perceived system effectiveness!
Hybrid system
They understand it
1! 1! **!
They are more TopN!
Sort!
satisfied with its UI -2! -1!
0!
0! 1! 2! Explicit!
They find it more Implicit!
Hybrid!
effective -1!
-2!
Domain Knowledge!
37. DOMAIN KNOWLEDGE
Experts like the ***!
2!
Perceived system effectiveness!
Hybrid system
They understand it
1! 1! **!
They are more TopN!
Sort!
satisfied with its UI -2! -1!
0!
0! 1! 2! Explicit!
They find it more Implicit!
Hybrid!
effective -1!
Control of Explicit -2!
and convenience of Domain Knowledge!
38. DOMAIN KNOWLEDGE
Experts like the
Hybrid system
They understand it
They are more
satisfied with its UI
They find it more
effective
Control of Explicit
and convenience of
40. DOMAIN KNOWLEDGE
Choices 2!
*!
1!
Experts make better
decisions with Choice satisfaction!
1! *!
Explicit, Implicit TopN!
and Hybrid 0!
Sort!
‣ But only Hybrid is -2! -1! 0! 1! 2! Explicit!
Implicit!
better overall
Hybrid!
-1!
-2!
Domain Knowledge!
41. DOMAIN KNOWLEDGE
Choices 2!
*!
1!
Experts make better
decisions with Choice satisfaction!
1! *!
Explicit, Implicit TopN!
and Hybrid 0!
Sort!
‣ But only Hybrid is -2! -1! 0! 1! 2! Explicit!
Implicit!
better overall
Hybrid!
Novices make better -1!
decisions with
TopN* -2!
‣ Unable to leverage Domain Knowledge!
42. DOMAIN KNOWLEDGE
Choices
Experts make better
decisions with
Explicit, Implicit
and Hybrid
‣ But only Hybrid is
better overall
Novices make better
decisions with
TopN*
‣ Unable to leverage
44. TRUSTING PROPENSITY
Trust is necessary to
accept the
recommendations
A lack of trust can cause reactance
Users have an initial trusting
propensity
Distrusting users may not
like Implicit
because they need a system that is
46. TRUSTING PROPENSITY
Distrusting 45!
users dislike
40!
User interface satisfaction!
35!
Explicit, 1!
Implicit, TopN
30! *! TopN!
25! Sort!
They are not 20! Explicit!
satisfied with the
Implicit!
UI
****! 15! Hybrid!
10!
5!
-2! -1! 0! 1! 2!
Trusting propensity!
47. TRUSTING PROPENSITY
Distrusting 2!
users dislike
Explicit, Perceived system effectiveness!
1!
Implicit, TopN TopN!
Sort!
0!
They are not -2! -1! 0! 1! 2! Explicit!
1!
satisfied with the
Implicit!
Hybrid!
UI -1!
They do not find *!
these systems *! -2!
effective Trusting propensity!
50. PERSISTENCE
Satisficers may like Implicit
The system updates the
recommendations to provide
similar items
Maximizers may like
Implicit or TopN
More counterfactual thinking, more
anticipated post-decision regret
This is aggravated in systems with
52. PERSISTENCE
2!
Choices **!
Maximizers are
1!
more satisfied with Choice satisfaction!
their choices! TopN!
Sort!
0!
-2! -1! 0! 1! 2! Explicit!
Implicit!
Hybrid!
-1!
-2!
Persistence!
53. PERSISTENCE
2!
Choices **!
Maximizers are
1!
more satisfied with Choice satisfaction!
their choices! TopN!
Sort!
Maximizers like -2! -1!
0!
0! 1! 2! Explicit!
their choices in Implicit!
TopN -1!
Hybrid!
-2!
Persistence!
54. PERSISTENCE
2!
Choices **!
Maximizers are
1!
more satisfied with Choice satisfaction!
their choices! TopN!
Sort!
Maximizers like -2! -1!
0!
0! 1! 2! Explicit!
their choices in Implicit!
TopN -1!
Hybrid!
Satisficers like
their choices in
-2!
Implicit* Persistence!
57. FIRST SOME RESERVATIONS...
Small sample of users
28-33 participants per condition; low power
Domain encourages multiple decisions; dampens
the effects
Results pertain to attribute-based
systems
Does not apply to collaborative filtering
58. CONCLUSIONS
Hybrid is better than Explicit and
Implicit
For experts: tweak preferences: convenience and
control
For distrusting users: negative reactions to other
systems
However, TopN may be better in some
cases
For novices: no knowledge to exploit the benefits
59. HOW TO COMBINE TOP-N AND
HYBRID?
Spatially separate them
In different sections of the interface
Temporally separate them
Start with the TopN, carefully introduce implicit
recommendations, then introduce explicit controls
Assign the correct method to each user
Discover the user’s characteristics,
then tailor the interface to her specific needs
60. MORE IN GENERAL...
Each to his own
The best interaction method depends on user
characteristics
Taking these into account may result in significantly
better recommender systems