Prity Khastgir IPR Strategic India Patent Attorney Amplify Innovation•24 views
Peer Ratings in Massive Online Social Networks
1. Peer Ratings in Massive Online Social
Networks
Dmitry Zinoviev
Suffolk University
Boston
Sunbelt-2012
2. SUNBELT-2012 / SUFFOLK UNIVERSITY 2
To Like or not to Like?
● Like/dislike ratings provide instant peer feedback in massive
online social networks (MOSNs)
● Objects subject to ratings: posts; photographs; comments
● Questions:
➢ How many stops should rating scales have?
➢ How do MOSN users perceive the ratings?
➢ How do MOSN users react to the ratings?
3. SUNBELT-2012 / SUFFOLK UNIVERSITY 3
Rating Types
● Single-valued (Facebook ”Like”, Google+ ”+1”); do
not allow negative attitudes
● Binary a.k.a. ”Thumbs Up/Down” (Pandora, Yahoo
Answers); do not allow fuzzy answers
● Multivalued (Yahoo, Amazon); may be hard to use
if there are many stops; these ratings are of
particular interest to us
4. SUNBELT-2012 / SUFFOLK UNIVERSITY 4
“Odnoklassniki.Ru”—Our Testing Range
● Largest Russian-language social network (100 mln users, 31 mln
daily visits)
● Most users have more than one personal picture and reasonably
credible demographics (age/gender)
● 6-way ratings (“1” through “5+”*) applicable to personal pictures
● Anyone can leave comments to any personal picture of to the profile
in general
● The system records and displays all profile visits (and visitors)
*The grade of “5+” requires a symbolic payment from the grader
5. SUNBELT-2012 / SUFFOLK UNIVERSITY 5
A Profile, as Seen by a Visitor
My PageMy Page
Personal PhotographsPersonal Photographs
DemographicsDemographics
FriendsFriends
Profile PictureProfile Picture
6. SUNBELT-2012 / SUFFOLK UNIVERSITY 6
Picture Rating Interface
PicturePicture
Picture CommentsPicture Comments
DemographicsDemographics
This is a ”5+” pictureThis is a ”5+” picture
Your ratingYour rating
DescriptionDescription
7. SUNBELT-2012 / SUFFOLK UNIVERSITY 7
Method
● Select a random avatar Q from the three avatars on the right
(30/Female, 40/Male, 30/Male)
● Select a random currently logged user and record the user's age
A (≥18) and gender G
● Select a random grade Rin
between “1” and “5+” and post it
● Record the user's response:
➢ did the user visited the test profile? (V)
➢ did the user post a comment? (C)
➢ what is the response grade Rout
?
●
Apologize to the user for a low grade. Posted grades and
comments can be easily removed without any traces
8. SUNBELT-2012 / SUFFOLK UNIVERSITY 8
Descriptive Statistics
● 10,799 observations
● 3,600 observations per avatar
● 1,800 observations per grade
● Distribution by gender:
➢ 54% female
➢ 46% male
● Mean age 34
Russia in 2009
9. SUNBELT-2012 / SUFFOLK UNIVERSITY 9
Empirical Parametrized Mapping
{Rout
, V, C} = f (Rin
; Q, A, G )
Free variable
Parameters
10. SUNBELT-2012 / SUFFOLK UNIVERSITY 10
Response Details with Trend Lines
Responders, by kind and by age:
11. SUNBELT-2012 / SUFFOLK UNIVERSITY 11
Response Statistics
All Responders
Of them:
Visitors Graders
Rate
M 47.0% 39.7% 19.7% 4.8%
F 42.4% 31.7% 19.5% 5.0%
M -0.1% 0.0% -0.4% -0.1%
F 0.5% 0.6% -0.1% 0.1%
Commenters
Age Trend
(%/year)
● Visits are the most common responses, followed by grades, followed
by comments
● Male subjects visit the experimenter's profile more often
● Older female visitors are more active than younger
● Younger male graders are more active than older
● Grading/commenting rates are gender-agnostic
12. SUNBELT-2012 / SUFFOLK UNIVERSITY 12
OutGrade Distribution
● The grades of “5” and “5+” are grouped (“5+” is special because of its
associated price)
● Response grades have a sharp bi-modal distribution
● Only grades “1” and “5” matter!
13. SUNBELT-2012 / SUFFOLK UNIVERSITY 13
OutGrades vs InGrades
● Response grades are sharply distributed around “1” and “5” for any
stimulus grade (the red lines show best-fit bi-exponential
distributions)
● Only two stops on the scale are necessary!
14. SUNBELT-2012 / SUFFOLK UNIVERSITY 14
RG=R+G
● Response Grade = Reciprocity + Generosity
➢ Reciprocity: Social norm of in-kind responses to the behavior of
others (a.k.a. “eye for an eye”)
Responding with the same grade
➢ Generosity: Habit of giving freely without expecting anything in
return
Responding with a better of worse grade (positive
vs negative generosity)
● How popular are these mechanisms?
15. SUNBELT-2012 / SUFFOLK UNIVERSITY 15
Reciprocity (1)
● Define reciprocity as the “fraction of reciprocally equal grades out of
all grades”:
● Calculate reciprocity for each avatar and for each age-gender group
● Calculate linear best-fit estimate
● Notation:
➢ Blue lines for male subjects, red lines for female subjects
➢ Solid lines for the 40/M avatar, dashed lines for 30/M, dotted lines
for 30/F
Rec=
N Rin=Rout
N
16. SUNBELT-2012 / SUFFOLK UNIVERSITY 16
Reciprocity (2)
● Older subjects
are less
reciprocating
● Male subjects are
on average less
reciprocating
● Younger avatars
generate more
reciprocity from
the subjects of
the same gender
17. SUNBELT-2012 / SUFFOLK UNIVERSITY 17
Generosity (1)
● Define generosity as the average value of the difference between the
stimulus grade and the response grade:
● It can be positive and negative
● Calculate generosity for each avatar and for each age-gender group
● Calculate linear best-fit estimate
Gen=∑
i
Rout−Ri n
N
18. SUNBELT-2012 / SUFFOLK UNIVERSITY 18
Generosity (2)
● Older subjects
are more
generous
● Older avatar
generates less
generosity
● Subjects are less
generous to the
avatars of the
same gender
● Younger avatars
generate less
generosity from
the subjects of
the same gender
19. SUNBELT-2012 / SUFFOLK UNIVERSITY 19
Benevolence (1)
● Let negative comments have the value of -1, positive comments—the
value of 1, and neutral comments—the value of 0. Define
benevolence as the average value all comments
● Calculate benevolence for each avatar and for each age-gender
group
● Calculate linear best-fit estimate
20. SUNBELT-2012 / SUFFOLK UNIVERSITY 20
Benevolence (2)
● Older subjects
leave better
comments
● Female subjects
leave better
comments
● Younger avatars
get worse
comments from
the subjects of
the same gender
21. SUNBELT-2012 / SUFFOLK UNIVERSITY 21
Grading/Commenting Overview (1)
Generosity G
Reciprocity R
Benevolence B
● Most behaviors are avatar-specific; however, the dependency on age
is universal
22. SUNBELT-2012 / SUFFOLK UNIVERSITY 22
Grading/Commenting Overview (2)
● Generosity and benevolence grow with age
● Generosity and benevolence are weaker for same-gender
avatar-subject pairs
● Reciprocity slowly declines with age
● Reciprocity is stronger for same-gender avatar-subject pairs
● All three values can be approximated using quadratic functions
23. SUNBELT-2012 / SUFFOLK UNIVERSITY 23
Latent Hostility
● Age 18–22: subjects are
all negative
● Age 22–36: subjects
combine positive
generosity and negative
benevolence: they post
higher grades but leave
negative comments
➢ Hypothesis: Higher
response grades make
subjects feel good, but
comments reveal true
feelings
● Age 36–80 subjects are all
positive
18–22
36–80
22–36