AND
IN
INSPECTABILITY
SOCIALRECOMMENDERS
CONTROL
ALEX BOSTANDJIEV
BART KNIJNENBURG
ALFR
ED KO
BSA
JOH
N O
’DO
NO
VAN
WEWHY SHOULD
SOCIALUSE
RECOMMENDERS?
DJ’S
VINYL?USE
WHY DO
BETTER INSPECTABILITY
INSPECTABILITY IN NORMAL RECOMMENDERS
BETTER INSPECTABILITYIN SOCIAL RECOMMENDERS
?THINGS I LIKE
VS.
RECOMMENDATIONS
THINGS I LIKE RECOMMENDATIONSFRIENDS
MAGIC
MORE CONTROL
CONTROLIN NORMAL RECOMMENDERS
MORE CONTROLIN SOCIAL RECOMMENDERS
VS.
RECOMMENDATIONS
RECOMMENDATIONS+
INTUITIVE INTERFACE
MORE CONTROL & INSPECTABILITY?MORE COMPLEXITY!
130 L. Chen, P. Pu
Step 4
Fig. 4 System showing a new set of alternatives after the user’s critiques
More specifically, after getting users’ initial preferences via the conversational dia-log, the system translates them into a SQL query and passes it to the database. If toomany matching goods exist, the Navigation by Asking function would calculate theinformation gain of possible questions and then ask appropriate questions to the shop-per for narrowing down the matching goods. After merchandise records are narroweddown to a pre-defined threshold number, the Navigation by Proposing function willshow three significantly different samples. The first sample good is the good recordclosest to the center point of all matching goods. Its selling points directly reflect thecustomer’s request. The second sample good is the record positioned farthest awayfrom the center point, and the third sample good is the one positioned farthest awayfrom the second sample. The explanation of the sample’s selling point is given like“this is twice as expensive as those because it is made of silk and the other two aremade of polyester”. While seeing the explanation, the shopper could easily excludeone of the three proposed goods with a specific reason, such as “this one is too darkfor me compared with the other two”. The system will then modify the sample pickingstrategy accordingly.
2.1.2 Adaptive place advisor (Thompson et al. 2004)
This system employs natural language dialog for giving personalized place recom-mendations. Consider the following conversation between the inquirer (i.e., the user)and the advisor (i.e., the system):
123
Critiquing-based recommenders 133
Fig. 5 The Dynamic Critiquing interface with system suggested compound critiques for users to select(McCarthy et al. 2005c)
2.2.2 MAUT-based compound critiques and visual critiquing (Zhang and Pu 2006)
However, the Dynamic-Critiquing method (including its extension) is still limited,in that it only reveals what the system can provide, but does not take into accountusers’ interest in the suggested critiques. Given this limitation, Zhang and Pu (2006)have proposed an approach with the purpose of adapting the generation of compoundcritiques to user preferences. Formally, they model each user’s preferences based onthe multi-Attribute Utility Theory (MAUT), which is a theory taking into accountof conflicting value preferences and producing a sore for each item to represent itsoverall satisfaction degree with the user preferences (Keeney and Raiffa 1976). Thus,during each recommendation cycle, according to the user’s current preferences, thetop k products with maximal MAUT values are retrieved. The ranked first item is thentaken as the top candidate, and for each of the others, its detailed value differencesfrom the top candidate will be presented as a compound critique (e.g., “same brandwith lower price, but slower CPU speed, smaller screen, smaller memory and smallerhard disk”).
Relative to Dynamic Critiquing methods, these MAUT-based compound critiqueswere proven with significantly higher recommendation quality, inferring that they can
123
144 L. Chen, P. Pu
(a) The preference-based organization interface.
(b) The user-initiated example critiquing interface.
If suggested critiques and products do not interest the user, he/she could switch to create critiques his/herself by clicking the button “Self specify criteria for ‘Better Features’.
Fig. 10 Hybrid critiquing system (version 2): the combination of preference-based organization interface(Pref-ORG) and user-initiated critiquing facility (Chen and Pu 2007b, 2010)
organization algorithm (as described in Sect. 2.2). If the user is interested in one ofthe suggested critiques, she could click “Show All” to see more products under thecritique. Among these products, the user can either choose one as her final choice, or
123
142 L. Chen, P. Pu
The product being critiqued
System-suggested compound critiques
User-initiated critiquing facility
Fig. 9 Hybrid critiquing system (version 1): the combination of system-suggested compound critiques anduser-initiated critiquing facility (Chen and Pu 2007a)
can not only obtain knowledge of the domain and easily perform critiquing via thesuggested critiques, but also have the opportunity to freely compose critiques on theirown with the self-initiated critiquing support.
3.2 Hybrid Critiquing
Two hybrid critiquing systems have been developed so far. One was the combinationof Example Critiquing facilities with Dynamic Critiquing based compound critiqueson a single screen. The second version integrated the preference-based organizationinterface (Chen and Pu 2007c) (which shows system-suggested critiques and theirassociated products on a separate page) with Example Critiquing (the EC interface isevoked only when users activate it). Series of user studies were conducted on the twoversions.
3.2.1 Version 1
Figure 9 shows a sample interface of the first type of hybrid critiquing system, whereDC compound critiques are displayed with EC facilities on the same screen (Chenand Pu 2007a). Specifically, the current recommendation is displayed at the top andfollowed by multiple suggested critiques. The self-initiated critiquing area is placedbelow, which provides functions to facilitate various types of critiquing modality
123
SIMPLECONTROL
SIMPLEINSPECTABILITY
THE POWER OF VISUALIZATION
AND MOREINSPECTABILITY
SATISFACTIONINCREASES
CONTROL
HYPOTHESIS:
BETTER
ONLINE USER EXPERIMENT
SYSTEM
Modified TasteWeights system
Facebook friends as recommenders
Music recommendations (based on “likes”)
Split up control + inspectability
PARTICIPANTS267 participants
Mechanical Turk + Craigslist
At least 5 music “likes” and overlap with at least 5 friends at least 10 recommendations
lists limited to 10 to avoid cognitive overload
Demographics similar to Facebook user population
PROCEDURESTEP 1: Log in to Facebook
System collects your music “likes”
System collects your friends’ music likes
PROCEDURESTEP 2: Control
3 conditions, between subjects
<skip>
NOTHING WEIGH ITEMS WEIGH FRIENDS
VS VS
PROCEDURESTEP 3: Inspection
2 conditions, between subjects
LIST ONLY FULL GRAPH
VS
6 CONDITIONS<skip> -->
-->
-->
<skip> -->
-->
-->
PROCEDURESTEP 4: Evaluation
For each recommendation:
Do you know this band/artist?
How do you rate this band/artist?(link to LastFM page for reference)
PROCEDURESTEP 5: Questionnaires
-understandability
-perceived control
-perceived recommendation quality
-system satisfaction
-music expertise
- trusting propensity
- familiarity with recommender systems
RESULTS
SUBJECTIVE3 items:- The recommendation
process is clear to me
- I understand how TasteWeights came up with the recommendations
- I am unsure how the recommendations were generated*
SUBJECTIVEINSPECTABILITY
full graph list only
CONTROL
SUBJECTIVE4 items:- I had limited control over
the way TasteWeights made recommen-dations*
- TasteWeights restricted me in my choice of music*
- Compared to how I normally get recommendations, TasteWeights was very limited*
- I would like to have more control over the recommendations*
full graph list only
SUBJECTIVE6 items:- I liked the artists/bands
recommended by the TasteWeights system
- The recommended artists/bands fitted my preference
- The recommended artists/bands were well chosen
- The recommended artists/bands were relevant
- TasteWeights recommen-ded too many bad artists/bands*
- I didn't like any of the recommended artists/bands*
full graph list only
SUBJECTIVE7 items:- I would recommend
TasteWeights to others
- TasteWeights is useless*
- TasteWeights makes me more aware of my choice options
- I can make better music choices with TasteWeights
- I can find better music using TasteWeights
- Using TasteWeights is a pleasant experience
- TasteWeights has no real benefit for me*
full graph list only
BEHAVIORTime (min:sec) taken in the inspection phase (step 3)
- Including LastFM visits
-Not including the control phase (step 2)
-Not including the evaluation phase (step 4)
full graph list only
BEHAVIORNumber of artists the participant claims she already knows
Why higher in the full graph condition?
- Link to friends reminds the user how she knows the artist
- Social conformance
full graph list only
BEHAVIORAverage rating of the 10 recommendations
-Lower when rating items than when rating friends
-Slightly higher in full graph condition
full graph list only
User Experience (EXP)Objective System Aspects (OSA)
Subjective System Aspects (SSA)
STRUCTURAL MODEL
User Experience (EXP)Objective System Aspects (OSA)
Subjective System Aspects (SSA)
+
Satisfaction with the system
(R2 = .696)0.410 (0.092)***
0.955 (0.148)***+
STRUCTURAL MODEL
+Perceived
recommendation quality
(R2 = .512)
0.770(0.094)***
+ Perceived control
(R2 = .311)0.377(0.074)***
+
+
Understandability
(R2 = .153)
Controlitem/friend vs. no control
Inspectabilityfull graph vs. list only
!2(2) = 10.70**item: 0.428 (0.207)*friend: 0.668 (0.206)**
0.459 (0.148)**
User Experience (EXP)Objective System Aspects (OSA)
Subjective System Aspects (SSA)
Interaction (INT)
+
Satisfaction with the system
(R2 = .696)0.410 (0.092)***
0.955 (0.148)***+
STRUCTURAL MODEL
Inspection time (min)(R2 = .092)
+
−
+
0.231(0.114)*
!2(2) = 10.81**item: −0.181 (0.097)1
friend: −0.389 (0.125)**
0.288 (0.091)**
Average rating(R2 = .508)
++
0.067 (0.022)**
0.323 (0.031)***
+
−
+
+
number of known recommendations
(R2 = .044)
−0.152 (0.063)*
0.249(0.049)***
0.695 (0.304)*
0.148(0.051)**
+Perceived
recommendation quality
(R2 = .512)
0.770(0.094)***
+ Perceived control
(R2 = .311)0.377(0.074)***
+
+
Understandability
(R2 = .153)
Controlitem/friend vs. no control
Inspectabilityfull graph vs. list only
!2(2) = 10.70**item: 0.428 (0.207)*friend: 0.668 (0.206)**
0.459 (0.148)**
User Experience (EXP)Objective System Aspects (OSA)
Subjective System Aspects (SSA)
Interaction (INT)
+
Satisfaction with the system
(R2 = .696)0.410 (0.092)***
0.955 (0.148)***+
STRUCTURAL MODEL
Inspection time (min)(R2 = .092)
+
−
+
0.231(0.114)*
!2(2) = 10.81**item: −0.181 (0.097)1
friend: −0.389 (0.125)**
0.288 (0.091)**
Average rating(R2 = .508)
++
0.067 (0.022)**
0.323 (0.031)***
+
−
+
+
number of known recommendations
(R2 = .044)
−0.152 (0.063)*
0.249(0.049)***
0.695 (0.304)*
0.148(0.051)**
+Perceived
recommendation quality
(R2 = .512)
0.770(0.094)***
+ Perceived control
(R2 = .311)0.377(0.074)***
+
+
Understandability
(R2 = .153)
Controlitem/friend vs. no control
Inspectabilityfull graph vs. list only
!2(2) = 10.70**item: 0.428 (0.207)*friend: 0.668 (0.206)**
0.459 (0.148)**
Personal Characteristics (PC)
+
−0.332 (0.088)***
0.205(0.100)*
0.375(0.094)***
+
−
Music expertise
0.257(0.124)*
+
Trusting propensity
0.166 (0.077)*
+
Familiarity with recommenders
CONCLUSION
CONCLUSIONSocial recommenders
- Give users inspectability and control
- Can be done with a simple user interface!
Inspectability:
- Graph increases understandability and perceived control
- Improves recognition of known recommendations
Control:
- Items control: higher novelty (fewer known recs)
- Friend control: higher accuracy
FUTURE WORKInspectability and control work
-Separately
-What about together?
FUTURE WORKInspectability and control work
-Separately
-What about together?
SOCIALRECOMMENDERS
LET YOUBE A
RECOMMENDATION
DJ
THANK YOU!WWW.USABART.NL
[email protected]@USABART
CONCLUSIONSocial recommenders
- Give users inspectability and control
- Can be done with a simple user interface!
Inspectability:
- Increases understandability and perceived control
- Improves recognition of known recommendations
Control:
- Friend control: higher accuracy
- Items control: higher novelty (fewer known recs)
full graph list only
User Experience (EXP)Objective System Aspects (OSA)
Subjective System Aspects (SSA)
Interaction (INT)
+
Satisfaction with the system
(R2 = .696)0.410 (0.092)***
0.955 (0.148)***+
STRUCTURAL MODEL
Inspection time (min)(R2 = .092)
+
−
+
0.231(0.114)*
!2(2) = 10.81**item: −0.181 (0.097)1
friend: −0.389 (0.125)**
0.288 (0.091)**
Average rating(R2 = .508)
++
0.067 (0.022)**
0.323 (0.031)***
+
−
+
+
number of known recommendations
(R2 = .044)
−0.152 (0.063)*
0.249(0.049)***
0.695 (0.304)*
0.148(0.051)**
+Perceived
recommendation quality
(R2 = .512)
0.770(0.094)***
+ Perceived control
(R2 = .311)0.377(0.074)***
+
+
Understandability
(R2 = .153)
Controlitem/friend vs. no control
Inspectabilityfull graph vs. list only
!2(2) = 10.70**item: 0.428 (0.207)*friend: 0.668 (0.206)**
0.459 (0.148)**
Personal Characteristics (PC)
+
−0.332 (0.088)***
0.205(0.100)*
0.375(0.094)***
+
−
Music expertise
0.257(0.124)*
+
Trusting propensity
0.166 (0.077)*
+
Familiarity with recommenders
Top Related