Mais conteúdo relacionado Performance and Scene3. PerformanceandScene
USER CLASSIFICATION PART 1/8
+ Humans have a hard time classifying information.
Advanced Naval/Air/
Technology Land battle
Space Travel Evil “Other”
Aliens Focus on
Focus on military life /
utopia / the effects of
apocalypse war
SCI-FI WAR
SCI-FI or WAR?
Brandon Webb, © 2012
4. *Can we
sort
people?
Brandon Webb, © 2012
6. + Example Natural Language Texts
+ product reviews (amazon)
+ status updates (facebook, twitter)
+blogs (tumblr)
+user profiles (OkCupid, A4A)
Brandon Webb, © 2012
7. +Language about
politics< concrete
than language about
products(Malouf& Mullen, 2008; Pang
& Lee, 2007).
+Hard to classify users
on political forums.
(Malouf& Mullen, 2008; Pang
& Lee, 2007).
WHAT ABOUT ON
ADAM4ADAM?
Brandon Webb, © 2012
8. is a public gaydating site.
Users create and edit
profiles,uploadpics,viewonline
members,and send one
another messages.
Brandon Webb, © 2012
11. + What is SCENE?
+ A4A allows 1-9 scene choices.
Brandon Webb, © 2012
12. + What is SCENE?
+ A4A allows 1-9 scene choices.
+ Individual identities are social
products (Jenkins, 2008).
Brandon Webb, © 2012
13. + What is SCENE?
+ A4A allows 1-9 scenes choices.
+ Individual identities are social
products (Jenkins, 2008).
+ Although two users share the
same nominal identity, their
lived experience may be
entirely different.
Brandon Webb, © 2012
15. PerformanceandScene
SCENEPART 3/8
Gay jock in downtown
Gay jock in rural
LA.
Nebraska
By laying claim to a scene, users communicate that they are more alikethan dislike.
Brandon Webb, © 2012
16. CASUAL
“I’d like to think that I’m a fairly unique person, but I guess everyone thinks
that about themselves, huh?”
Brandon Webb, © 2012
17. TRENDY
“Chill guy here…love photogrphy, art galleries, read a good book in a coffee
chop, indie films, fashion magz and clubs.”
Brandon Webb, © 2012
18. ALTERNATIVE
“I like to chill with people who are conscious with their world and other people’s
worlds. I like people who can carry good conversations and be open minded.”
Brandon Webb, © 2012
19. MILITARY
“GL laid back military dude. Into other built guys but not lookin for random
hookups.”
Brandon Webb, © 2012
20. JOCK
“Looking for sexy athletic hot bodied men for pleasurable times. If you are not
athletic we are probably not a good match.”
Brandon Webb, © 2012
21. DRAG
“Skater dude by day rock Princess by night. yes I am a drag queen, but don’t let
that scare you off it pays the bills goes to show how fun I can be…”
Brandon Webb, © 2012
23. PUNK
“Laid back sort of guy. Like other men with similar interests. Not into fem
guys at all. Like music and have a soft spot for punk style guys…”
Brandon Webb, © 2012
24. CONSERVATIVE
“Regular guy into sports, gym a couple times a week. neg/std free, no drugs,
play safe only, looking for similar guys for discreet times or hanging out…”
Brandon Webb, © 2012
30. PerformanceandScene
PERFORMANCEPART 4/8 Navy guy here,
be discreet Athletic guy looking for
other dl dudes
GI bi discreet milguy here Really
into other military or jock guys
MATCH
USER Brandon Webb, © 2012
31. PerformanceandScene
PERFORMANCEPART 4/8
Anybody down for coffee or a
PUBLIC movie? Regular guy here love
the outdoors!
Hey boys, I love hiking and Hey guys the names Rick I’m a
basketball music and dance double major…
Hey just looking to make
some friends Sup!
Brandon Webb, © 2012
34. PerformanceandScene
CORPUSPART 5/8
+ 2,437 user profiles were taken from A4A (IRB APPROVAL: 777080).
+ Each user profile stored in XML file tagged for performance and
scene = this became the training corpus.
<USER ID=001>
<PERFORMANCE>
”Looking for NSA fun. NOT looking for a
relationship, already have that”
</PERFORMANCE>
<SCENE>
Casual
</SCENE>
</USER>
…
<USER ID=2437>
…
Brandon Webb, © 2012
35. PerformanceandScene
CORPUSPART 5/8
+ Composition of training corpus
499jock
109military
238alternative
577casual
399trendy
41punk
462conservative
92leather
20drag
Brandon Webb, © 2012
37. PerformanceandScene
NAÏVE-BAYESPART 6/8
+ How do we classify users?
+ Each performance is transformed into a list of features so that it
can be read by the classifier
“Looking for NSA fun. NOT looking for a relationship, already have that.
Versatile, but prefer bottom. NO tweekers or PNP losers, been there,
was one.”
becomes a list of features with functional language deleted…
*‘looking’, ‘nsa’, ‘fun’, ‘looking’, ‘relationship’, ‘already’, ‘have’, ‘versatile’,
‘prefer’, ‘bottom’, ‘tweekers’, ‘pnp’, ‘losers’, ‘been’, ‘there’, ‘was’, ‘one’+
Brandon Webb, © 2012
39. PerformanceandScene
NAÏVE-BAYESPART 6/8
+ We do this for all 2,437 profiles so that we end up with a huge list
of features.
+ Then, for each performance, what features do we witness in that
performance that we also witness in the list of features? If a
performance has a feature, we mark it with 1. If not, 0.
(,contains(‘gear’): 1, contains(‘workout’): 0, contains(‘sup’): 1, …, -, jock)
(,contains(‘leather’): 1, contains(‘workout’): 0, contains(‘sup’): 0, …, -, leather)
+ In this way, each performance becomes a vectorthat we feed
to the Naïve-Bayes classifier for training.
Brandon Webb, © 2012
40. PerformanceandScene
NAÏVE-BAYESPART 6/8
NAÏVE-BAYES TRAINING PHASE PSEUDOCODE
TRAIN_NBCLASSIFIER (PERFORMANCE, SCENE):
1 F← EXTRACT_ALL_FEATURES(PERFORMANCE, SCENE)
2 V ← TRANSFORM_INTO_VECTOR(PERFORMANCE)
3 For each scene:
4 DO NS← COUNT_SCENE_OCCURENCES(SCENE)
5 PRIOR[S] ← NS/N
6 For each vector in V:
7 For each feature in F:
8 Tscene+feature← COUNT_FEATURE_OCCURENCES(VECTOR)
9 DO CONDPROB[scene][feature]← Tscene+feature/ Σfeature’ * (Tscene+feature’)
10 RETURN (prior, condprob)
Brandon Webb, © 2012
41. PerformanceandScene
NAÏVE-BAYESPART 6/8
NAÏVE-BAYES TESTING PHASE PSEUDOCODE
CLASSIFY_USER(PRIOR, CONDPROB, PERFORMANCE):
1 W←EXTRACT_FEATURES(PERFORMANCE)
2 For each scene:
3 DO SCORE[C] ← log(PRIOR[S])
4 For each feature in W:
5 DO SCORE[S]+= log(CONDPROB[feature][scene])
6 RETURN ARGMAX(SCORE[S])
Brandon Webb, © 2012
42. leather
ltr dom hiv
handcuffs
pig
bondage
hiking
btmwstoys gear
masculine
daddy
wild
guy
dominant hairy
workout
fems
couples
sex regular
snowboarding
poz explore
dance
pnp
mild
holes
sub
bears
fats
workout
water
Brandon Webb, © 2012
43. So how do users performsocial identity?
Brandon Webb, © 2012
44. PerformanceandScene
RESULTSPART 7/8
Naïve-Bayes classifier results (scene vs. scene and multi-scene)
Scene CAS TRE ALT MIL JOC LTH PNK DRG CNS
CAS - .63 .55 .50 .63 .53 .50 .50 .75
TRE .63 - .60 .50 .73 .53 .50 .50 .78
ALT .55 .60 - .55 .63 .53 .50 .50 .63
MIL .50 .50 .50 - .50 .73 .50 .50 .50
JOC .63 .73 .63 .50 - .53 .50 .50 .73
LTH .53 .53 .53 .73 .53 - .50 .53 .53
PNK .50 .50 .50 .50 .50 .50 - .55 .50
DRG .50 .50 .50 .50 .50 .53 .55 - .50
CNS .75 .78 .63 .50 .73 .53 .50 .50 -
ALL .217
Brandon Webb, © 2012
45. CRAZY
RETAIL ARTIST
FASHION
MOVING
MOVED
SHOPPING
THEATRE
BEFORE
FLOW
SHY HOLD
SWEET
Brandon Webb, © 2012
46. CRAZY
NASTY ARTIST
SWEATYFUCKIN
CREATIVE
SINGER
TATTOOS
INTENSE
LATE
BLAH
BAR POZ
THIN
Brandon Webb, © 2012
48. CRAZY
TRADE SECURE
DIVING MILITARY
NAVY
PARKS
FETISH
SNOWBOARDING
HOLE
YO
GOVT ARMY
PIX
Brandon Webb, © 2012
49. SPORTS
TRADE SUP
BASKETBALL LIFTING
CURIOUS
JOCK
BUDS
GEAR
SUM STD PIX
Brandon Webb, © 2012
50. DICKS
ANYWAY
LICKED VALUES
MAKEUP
FEM MOVING
MOVED
SHOPPING
FILLED
CLUBS
WEEKEND
CALI CD TX
Brandon Webb, © 2012
51. HOLES
DOM
EXPLORE
WILD COUPLES
BONDAGE
PIG LEATHER
BEARS
MILD
TOYS
FILLED
LTR
SUB WATER POZ
BTM
Brandon Webb, © 2012
52. TEND
FASHION
ODD PIERCED
SERIOUSLY
ARTLAIDBACK
LATINOS
MOMENTS
PIERCED
LTR
WHITES
WANNA
FOOL
HATE
Brandon Webb, © 2012
53. PerformanceandScene
DISCUSSIONPART 8/8
+ So what do we know about how users perform social identity?
+ Users sculpt similarity by appealing to matches and dissimilarity
by dissociating from trolls.
I’m a well put together & I’m not into GROSS,
adjusted 40’s guy. Fun, GAMES, LIARS,
educated, Driven, Self CHEATERS/PARTNERED
Employed Professional, MEN, or ANONYMOUS
Private, Quiet, Focused & SEX.
Healthy.
MORE LIKE THESE LESS LIKE THESE
MATCHES TROLLS
Brandon Webb, © 2012
54. PerformanceandScene
DISCUSSIONPART 8/8
+ So what do we know about how users perform social identity?
+ Users sculpt similarity by appealing to matches and dissimilarity
by dissociating from trolls.
+ Users sculpt identitythrough references to public and private
behavior.
MORE LIKE THESE LESS LIKE THESE
MATCHES TROLLS
Brandon Webb, © 2012
55. PerformanceandScene
DISCUSSIONPART 8/8 MORE LIKE THESE
MATCHES
+ Jocks sculpt similarity through appeals to ACTIVE LIFESTYLES, SPORTS,
GYM, FITNESS, and VERNACULARin a public ceremony.
Me: 6’ tall, 180 lb, 6% body fat.
Loves working out anything
I work out 5x a week that challenges me mentally
an am a pretty in and physically.
shape guy
Wuz up . . . Lifting
&blading
Brandon Webb, © 2012
56. PerformanceandScene
DISCUSSIONPART 8/8 MORE LIKE THESE
MATCHES
+ Leather users sculpt similarity through appeals to NONSTANDARD
LIFESTYLES, KINK SEX, andSUBCULTURAL PARAPHERNALIAin a private
ceremony.
Available to please a
guy, worship his body, Feel free to spit on me or
following orders. verbally insult me.
Bondage, hoods, gags, nipplepl
ay.
Looking for eager or
curious subs and
doms.
Brandon Webb, © 2012
57. PerformanceandScene
DISCUSSIONPART 8/8
+ When a performance displays behavior that is contradictory of
the scene as a category, it is difficult to classify that user.
Normal guy, going to work Love gangbangs, anon
every day and enjoying life. (blindfolded cool), pump and
Love to travel and create dump style fucking. Can get
adventures. What’s up with wild also, like ws and light
people flipping off others as bondage
their primary pic?
Brandon Webb, © 2012
58. PerformanceandScene
DISCUSSIONPART 8/8
+ Jocks use VERNACULAR to appeal to matches, whileleather users (if
at all) use VERNACULAR to dissociate from trolls.
“Don’t sup me, or you will be ignored.”
“Sup guys, chill dude here lookin to
chat.”
Brandon Webb, © 2012
59. PerformanceandScene
DISCUSSIONPART 8/8
+ Jocks use VERNACULAR to appeal to matches, whileleather users (if
at all) use VERNACULAR to dissociate from trolls.
“Don’t sup me, or you will be ignored.”
“Sup guys, chill dude here lookin to
chat.”
+ Due to stereotypes which suggest leather-identified men are
unorthodox, some attempt to repair their image through appeals
to ACTIVE/STABLE LIFESTYLES and EMPLOYMENT.
“I am basically blue collar who works very hard.”
“Easy goin, outdoors camp kayak hike bike.”
Brandon Webb, © 2012
60. PerformanceandScene
DISCUSSIONPART 8/8
+ Jocks use VERNACULAR to appeal to matches, whileleather users (if
at all) use VERNACULAR to dissociate from trolls.
“Don’t sup me, or you will be ignored.”
“Sup guys, chill dude here lookin to
chat.”
+ Due to stereotypes which suggest leather-identified men are
unorthodox, some attempt to repair their image through appeals
to ACTIVE/STABLE LIFESTYLES and EMPLOYMENT.
“I am basically blue collar who works very hard.”
“Easy goin, outdoors camp kayak hike bike.”
+ Jocks more often than other scenesdissociate from trolls
through references to BODY TYPE.
“ONLY into guys who r in good shape”
“Please be fit if not we probably wont work.”
Brandon Webb, © 2012
62. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
Brandon Webb, © 2012
63. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
+ There are regularities and systematicities to performance that the
classifier missed.
Brandon Webb, © 2012
64. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
+ There are regularities and systematicities to performance that the
classifier missed.
+ Training on a larger corpus may improve results.
Brandon Webb, © 2012
65. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
+ There are regularities and systematicities to performance that the
classifier missed.
+ Training on a larger corpus may improve results.
+ Tagging each performance for shifts in target audience may
improve results.
Brandon Webb, © 2012
66. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
+ There are regularities and systematicities to performance that the
classifier missed.
+ Training on a larger corpus may improve results.
+ Tagging each performance for shifts in target audience may
improve results.
+ Tagging each performance for shifts in discussion of public or
private behavior may improve results.
Brandon Webb, © 2012
67. PerformanceandScene
CONCLUSIONFINALE
+ Tried to classify users on A4A based on performance.
+ Naïve-Bayes classifier returned okay results, but failed to capture
the subtle complexities of discourse structure.
+ There are regularities and systematicities to performance that the
classifier missed.
+ Training on a larger corpus may improve results.
+ Tagging each performance for shifts in target audience may
improve results.
+ Tagging each performance for shifts in discussion of public or
private behavior may improve results.
+ Or use other classifiers, like Support Vector Machines.
Brandon Webb, © 2012