graphs 101
i go[e
1
node
node
sodas networks (facebook): nodes are people, edges friendship
communication graph (skype): nodes are people. edges communications
search ranking graph (google): nodes are pages. edges links
taste graph (hunch): nodes are people. edges taste similarity
edge
interest graph (twitter, instagram): nodes are people, edges interest
graphs 101
social networks (facebook): nodes are people, edges lriendshipe
communication graph (skype): nodes are people. edges communications
search ranking graph (google): nodes are pages. edges links
taste graph (hunch): nodes are people. edges taste similarity
node
edge
node
interest graph (twitter, instagram): nodes are people, edges interest
graphs 101
node ___________ node
edge
social networks (facebook): nodes are people. edges lriendshipe
communication graph (skype): nodes are people, edges communications
search ranking graph (google): nodes are pages, edges links
taste graph (hunch): nodes are people. edges taste simharity
interest graph (twitter, instagram): nodes are people, edges interest
first graph theory:
euler’s 7 bridges of koeningsberg
‘is it possibe to traverse the town & cross each brkge exty once?
convert ‘and to nos & bikis to edges
•any node at psed through mu have ev iiimber ci edges
‘thus xdy sdvate it you have 0 2 no wth odd mirr of edges
first graph theory:
euler’s 7 bridges of koeningsberg
ispossibeto baversethetown &as eh idge ectty ?
cwert land to nodes & bñdges to edges
•any node that is passed tirnxigfl mu have even nimiber of edges
‘thus y solvable if you have 0 or 2 no with odd num of edges
first graph theory:
euler’s 7 bridges of koeningsberg
convect iam to nodes & *ges to es any node th s psed thn4i must have even number edges
•ttwis cdy svabse i(yj have 0 2 nodes with odd number c edges
he
le 7
goie
1
.45 t posseie to traverse the town &crs each bridge exactty once?
undirected graph: relationship symmetric
(friendship)
il
• c a . —  i—• •
—
i
4
j _• i__r
— r,._——.  
__a .
cttkc ‘—

interlude: data fun
averages
twitter
number of followers: 62.97 per user
number of followees: 43.52 per user
facebook:
number of facebook hkes: 217.2 per item (liked)
number of facebook likes: 29.30 per user
but distributions are interestingly differenl.
—
1—
—
averages
twitter
number of followers: 62.97 per user number of followees: 43.52 per user
facebook:
number of facebook likes: 2172 per item (liked)
number of facebook likes: 29.30 per user
but distributions are interestingly different..
averages
twitter
number of followers: 62.97 per user
number of followees: 4152 per user
facebook:
number of facebook likes: 217.2 per item (liked)
number of facebook likes: 29.30 per user
but distributions are interestingly different..
averages
twitter
number of followers: 6197 per user
number of followees: 43.52 per user
facebook:
number of facebook likes: 2172 per item (liked)
number of facebook likes: 29.30 per user
but distributions are interestingly different..
i’il
twitter distributions are power curves
distribution oft of fdllowec5 you have thbn of # of people you follow
twitter distributions are power curves
distribution of # of followers you have
distribution of i of people you follow
a——
l
spike of # followinge curve around 20 due to old onboarding process (?)
twitter distributions are power curves
distribution of i of followers you have
i go
li
distribution of i of people you follow
a—
—p
s a. — — w 5— ws s — rn 
p 8 p 0 * ? • w
spike of # foiiowing’ curve around 20 due to old onboardirig process (?)
twitter distributions are power curves
distribution of s of followers you distribution of # of people you follow
spike of # following curve around 20 due to old onboarding process (?)
twitter distributions are power curves
distribution of i at followes you have
distribution of i of people you follow
w •
spike of # foilowing’ curve around 20 due to old onboarding process (?)
twitter distributions are power curves
distribution of # of followers you have
distribution of # of op1e you follow
spike of # foilowing curve around 20 due to old onboarding process (?)
‘i
1 go1e
facebook friends is more like a bell curve
i
s 9 13 17 21 x) 37 41 ii 1 57 61 73 77 si 93 97
y = number of people; x = number of friends for those people
i
facebook friends is more like a bell curve
1 5 9 13 17 21 37 *1 49 57 61 73 77 si 97
y = number of people: x = number of ffiends for those people
i
facebook friends is more like a bell curve
i
¶ 5 9 13 it 21 fl *1 49 1 57 61 73 77 81 v
y = number of people; x = number of ffiends for those people
of f&aoo’i l.s
0
facebook “likes” similar to twitter (since
also nonsymmetric?)
, 5 , ?3 17 21 25 29 fl v 41 .l5 53 57 61 6 73 77 31 1% 33 97
f’
t v.
purchased product
marketing
• — = tcs
tii
ll
!go1e
i
similar demographics to a
4. _._e_.._.._.
communicates with a b more likely to buy than c
telecom company tested using phone call graph to use for direct mail
targeting network neighbors of purchasers dominated other targeting techniques.
today, facebook and many ad networks use similar targeting for online ads
—  *  shawnar.i 4t i foster
marketing
similar demographics to a
4.....
purchased product communicates with a
b more likely to buy than c
telecom company tested using phone call graph to use for direct mail
targeting network neighbors of purchasers dominated other targeting techniques
today, facebook arid many ad networks use similar targeting for online ads
—: — “  sh.awra4t1. fser
prcs an crs .c:’s
d
_pi
/
4
i 
___ go gle
therefore
4
henchman
a
defense
you can infer organizational hierarchies from communication patterns. governments use this to map rogue organizations.
calls
responds immediately
calls
4.. . . .
responds slowly
boss
google founders’ $200b idea wocds and documents are nodes. connected by occurrence
pagerank: links are directed graph node
node
;1] yjr ws eme’
— l s9;0]
—•_:
.
__ — a. — a ——
• zam•
— l
arr crs te ov) ‘_...., t lr. p..
—
google founders’ $200b idea
words and documents are nodes. connected by occurrence
pagerank: links are directed graph
—  
node
;1] __ ___ y, en.r
—;0]
— — • —— —
node
t. :(
— z
t,çan ay to ci p trz . p._ —

r.
‘
li
,
google founders’ $200b idea words and documents are nodes, connected by occurrence
pagerank: links are dwected graph
node node
_______
 ____ __ __
e1li. army to ck
t.r.j’_1—i. p i—i
;1] ‘yql ne 5 ett
 — ;0]
— 
—
— —.

— p —
google founders’ $200b idea
words and documents are nodes, connected by occurrence
pagerank: links are directed graph
node node
“— _____

.
—4t to cvàh •j ‘r.
—
zinggraphs
start with smaller graph:
bowling pin strategy
• utility is proportional to square of network coverage, but how to start?
• shnnk size of me initial network arid grow from there
• also try to choose a subnetwork with natural ‘spillover’ effects
•ln this example, students at one college tend to have friends at others
i
n
>
start with smaller graph:
bowling pin strategy
• utility is proportional to square of network coverage, but how to start?
• shrink size of the initial network and grow from there
• also try to choose a subnetwork with natural ‘spillover’ effects
•f n this example, students at one college tend to have friends at others
a
0
a
h.
—
find clusters within existing graphs
i’ll;
,
a lot of people in the 90s thought dating would be winner
take all  but didn’t account for clustered graph structure
lf
go ‘gte
introducing overlap of buyers/sellers can add differentiation even in entrenched graphs
heterogeneous homogenous
hyt)ñd buyers/sellers
openlable
li*:
aii
_rn4s li
for heterogenous buyers/sellers consider ladies night strategy
introducing overlap of buyers/sellers can add
differentiation even in entrenched graphs
heterogeneous buyers/sellers
hybñd
homogenous buyers/sellers
opentable
• harmony’
p
4word )
linked
[ 
for heterogenous buyers/sellers consider ladies night strategy”
introducing overlap of buyers/sellers can add differentiation even in entrenched graphs
heterogeneous homogenous
buyers/sellers hybnd
opentable
li
aw
()
for heterogenous buyerslsellers consider ladies night 5frateg
introducing overlap of buyers/sellers can add differentiation even in entrenched graphs
heterogeneous buyers/sellers
0 pentabie
hybrid
homogenous buyers/sellers
i’ 
— —
for heterogenous buyers/sellers consider ladies night strategy”
7tioflstqr
7 f.;: i

ic
v
introducing overlap of buyers/sellers can add
differentiation even in entrenched graphs
heterogeneous homogenous
buyers/sellers hy buyers/sellers

openiable
“.3
li
;1] gaycom;0]
for heterogenous buyers/sellers consider ladies night strategy”
introducing overlap of buyers/sellers can add
differentiation even in entrenched graphs
heterogeneous homogenous
buyers/sellers hybnd
opintable
l:’*
..___j —w
il)
;1] gaym;0]
for heterogenous buyerslsellers consider ‘ladies night strategy
when to interoperate?
metcalfe’s law
network vakie  (nodesj
corollary
litite guy benefits ne than big guy
le jy joins rwtwork and:
‘big guy gáns smal ienntaj inaease m cnetkws ‘le guy garns vaàue & the rriy ex!sting conneis
•ths wtiy aim (as irintent big pay) r?sd wnen yahc! & googe vard to int’operate for im
big guy
little guy
when to interoperate?
metcalfe’s law
netwcxk varie — (nod
corollary.
littse guy benefit more than big gu
big guy
little guy
løe :j
rita ncre.sr
l:t1egy 3
•—‘ •jwd  &gvoew3kt cqecate’c’
when to interoperate?
corotiary
liwe guy benefits cre than bq guy big guy
l giqs ret1, 4
rea
.‘: ‘‘: d!
& gooe warte z ‘ ercçate fr m
metcalfe’s law netwark vakie  (noiies)
little guy
0
v
when to intemperate?
metcawe’s law
network vamie  (nos
corollary:
little guy benefits nre thaws big guy
l rvetwcrk awi .e rrai rease
—  ‘:‘ ?jm  
&gooewac,c.
big guy
little guy
i
when to interoperate?
big guy guy
ui’ iyrei a
6k! . rceas
e j1hs uethen3, ‘“:
•thas .‘y um cis t’’
& gooqe .yar!e2 :d .
i
jii
go ‘gle
metcalfe’s law
netwk ahie  (riod
1
corollary:
uttie guy benets nm tin b guy
1
when to interoperate?
metcalfe’s law
netwtxk va1i  nocs)3
corollary:
lime guy beneñts me thai bg guy
‘a
l* w ic retwc’ art
“bir y : t3i ;rmetaj encease
te guy au’.s ; e —arv
‘ •‘: ,j . 
& gooqe warte
go g1e
bg gur
little guy

on the other hande..
i —. • —
1 —
_____ facebook dev platform
p
r
eacn liffle guy benefits more than the big guy from interoperating
butthousands of little guys relying on the big gu lidifies big guy position
• facebook reaiized this and introduced facebook apps. connect and other ainteroperatinge features to prevent the social network deca’ that destroyed previous social networks.
j __
____z
on the other hand...
• each little guy benefits more than the big guy from interoperating
• but thousands of little guys relying on the big guy solidifies big guy position
• facebook realized this and introduced facebook apps. connect and other intemp&ating features to prevent the ‘social network decay that destroyed previous social networks.
/
v
facebook dev platform
/
shameless seifpromotion: taste graphs
i i
tastemates as basis of a graph
someone out there must enjoy the same tile/strategy games i do...
and chances are they are not (yet, anyway) my friend
enigmo

modem conflict
i
‘e
i
__ 4
1.
.9
(.arcasonne
tastemates as basis of a graph
someone out there must enjoy the same tilelstrategy games i do..
and chances are they are not (yet anyway) my ffiend
j
modem conflict
 ç
i
 w; _tl
enigmo
a,
4
r
(arcasonne
tastemates as basis of a graph
someone out there must enjoy the same tile/strategy games i do....
and chances are they are not (yet anyway) my friend
l.1
i 1
l
p.
i
enigmo
moiem conflict
carcasonn e
the “cold starr challenge for tastebased predictions
p
how to provide initial recommendations for a new user?
go ‘g1e _____________
ii netfiik forcetrainthenpredict
facebook metastesñvenbysociaigraph
kleverage crossvertical knowledge and
u flc adjacent known nodes in taste graph
the “cold start challenge for tastebased predictions
n f t f [i force train, then preiict
facebook assume tastes are driven by social graph
kleverage crossvertical knowledge and
u flc adjacent known nodes in taste graph
the “cold start” challenge for taste—based predictions
how to proviae initial recommendations for a new user?
netf[1 forcetrainthenpredict
;1] ___
facebook;0]
hunch
assume tastes are driven by social graph
everage crossverticai nowieage ano adjacent known nodes in taste graph
the “cold start” challenge for tastebased predictions
itow to provide initial recommendations for a new user?
force train, then predict
facebook
assume tastes are driven by social graph
leverage crossvertical knowledge and adjacent known nodes in taste graph
the starts challenge for taste—based predictions
how to prowle initai recommengatlons for a ew user’
force train, then predict
r cebook assume es are drtven by social graph
everage crossverucai riowieage ano
u fl c ajacent known nodes in taste graph
the “cold start” challenge for tastebased predictions
how to provide initial recommendations for a new user?
n ft f [i x force train. then predict
assume tastes are driven by social graph
beverage crossverticat nowteage ana
h u fl c h adjacent known nodes in taste graph
one cold start solution:
propagate known data to unknown nodes
• iteratively propagate with adjacent data
• dynamically adjust with ‘hard’ data
• lath er rinse, repeat
known data 0= unknown data
i
4
.
fun with apis
youzakk. automaticdj
1_— —
i
ri
benjamin gtettzman
• —4 \_.
i
applications
gog1e
fbi., i.
3
—
thetflix predictions for everything
ecommerce and mobile
[1
i—


b
fun with apis
youzakk. automaticdj
applications
i
u
i i1i
go ‘gle
_i4
aofletfijx predidions for everything
ecornmerce
and mobile
..tr
—
lp_

l.
njamin gli
applications
afletfljx predictions for everything
llei
f
f : i 3i’*a
f
—
fun with apis
youzakk, automaticdj
liz
_‘‘i_i
j.
ecommerce
and mobile
all
i
—
mtn gan
netflix predictions for everything
ecommerce
and mobile
applications
— ,

i—.—.’ —
i. _!
in
fun with apis
youzakk. automaticdj
benjamin gleitzmari
‘4 i
‘netffix predictions for everything
applications
ecornmerce
and mobile
—
t.__— —p..,,
[.j:
ti .
— —
._!
qi
i. ‘:=i.
1
, iil
fun with apis
youzakk. automaticdj
ijaniin glej7il
communications graphs:
how related are they to social or taste graphs?
______ my iphone contacts include some of my ñiend&
__ but also my plumber, doctor, network admnistrator united
airlines and the chinese restaurant around the corner
— ‘ a — — —
a lot of people were surprised that their email contacts were
assumed to be active social contacts
: 
i i
could we use ad preferences to cold start restaurant recs?
got igle hotpot
+
google _
— — — —— — — ,
f.
—
& i . t  ____
____
a—s, a n a r 
e ._ .‘. r _____
*rl
— a ‘ a...a.— — ____
____• — a t  _____
a •  a —
 i. ____
ti ____
i •  _____
____  i   _____
tr ‘ é cz ____
. t caa. i.. a — _____
“—f — — —i
we know this person likes classical music, yoga. poetry, and hiking
goügle
 — — —‘ ‘——
‘.

i
ai...
rr4p
i  °‘ i
j
ws
* i shr.. ij .
• — a c.
 t tz a
: — • r
l i 3r
,a.——’_
 1. — .,l a
34
cross domain data can solve the “napoleon
—— —— —
dynamite” problem
— —.. —— _ii.
i’••
i’
a — a
a
1

e. —
—
si
go ‘gle
— — _,._i_____ z
a — a
i — 