U C L

1
daniele quercia
U C L
2
I finished my PhD _at_
3
University College London
4
ltMy Researchgt
5
Ratings on ...
6
Ratings on phones
7
Why ratings on mobiles?
8
Daniele Quercia
Situation People exchange
digital content
9
People
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
The problem is ...
14
drowning user (content overload)
help!
who will come to the rescue?
15
Proposal Accept content only
from reputable people
16
Use
Store
17
MobiRate
Use
Store
Ubicomp08
18
LDTP
MobiRate
Use
Store
ICDM07
Ubicomp08
19
How to store ratings?
20
2. Gossip(to check each credential)
1.Log(credentials)
21
2. Gossip(to check each credential)
1.Log(credentials)
?
?Impractical
?
22

Idea behind MobiRate
23
Lets make it practical...
24
1.Sealed Log(of credentials)
2. Gossip(to check seals only)
25
1.Sealed Log(of credentials)
2. Gossip(to check seals only)
?
?Practical
?
26
1.Sealed Log(of credentials)
entry (rating)
seal (for the entry)
27
1.Sealed Log(of credentials)
entry (rating)
seal (for the entry)
hash chain binding whole table
28
Assumption ID is a unique public key
29
witnesses
slow down
30
What witnesses do
Audit!
slow down
31
What witnesses do
Audit!
Why?
slow down
32
(No Transcript)
33
Exposed
34
Who are my witnesses
Those who will share content with me
35
.
Who are my witnesses
Like-minded familiar strangers
36
1.Sealed Log(of credentials)
2. Gossip(to check seals only)
?
?Practical
?
37
Does MobiRate work?
38
Mobility Traces AND Social Networks
Reality Mining
39
Does MobiRate work?
lt1gt Is it robust to malicious individuals? lt2gt
Does it run on phones?
40
lt1gt robust
The f factor
41
(No Transcript)
42
(No Transcript)
43
Oracle
MobiRate
44
MobiRate reduces f!!!
45
lt2gt run
46
heaviest protocol runs lt 2sec
47
longest protocol completed
in 2.5ms (if Bluetooth 100kb/s)
48
MobiRate works
? robust ? runs on phones
49
MobiRate
Use
Store
Ubicomp08
50
LDTP
MobiRate
Use
Store
ICDM07
Ubicomp08
51
Use ratings? To make predictions!
52
Daniele Quercia
Traditional way Trust propagation
C
?
A
B
53
Daniele Quercia

• That way works on
• Web binary ratings

54
(No Transcript)
55
(No Transcript)
56
(No Transcript)
57
(No Transcript)
58
(No Transcript)
59
Daniele Quercia
60
Daniele Quercia
61
Daniele Quercia
62
Daniele Quercia
graph ? loss(error) ? function(min error)
63

Idea behind LDTP
ICDM07 Lightweight Distributed Trust Propagation
64
Daniele Quercia
C
1
2
?
A
B
65
Daniele Quercia
new graph
?
C
1
2
?
A
B
66
Daniele Quercia
new graph
good rating function
?
C
1
f
2
?
A
B
67
A ? B
68
A ? C
A ? B
A ? D
C ? B
69
A ? C
A ? B
A ? D
C ? B
group ratings
70


1)The relationships with same rater (A)
A
B
1
2
3
2
D
C
1
As long as B and D behave alike
71

2)The relationships with same rated node(B)
A
B
1
2
3
2
D
C
1
As long as A and C rate alike
72
A ? B
73
A ? C
1
A ? B
0.5
1
A ? D
C ? B
74
A ? C
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
75
Daniele Quercia
? new graph
good rating function
?
C
1
f
2
?
A
B
76

• Take f that returns
• the ratings already there ( )
• similar ratings for neighbouring nodes

A ? C
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
2
3
77

• Take f that returns
• the ratings already there ( )
• similar ratings for neighbouring nodes

A ? C
L(f)? ? M (f(A?D)-2)2
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
2
3
78

• Take f that returns
• the ratings already there ( )
• similar ratings for neighbouring nodes

A ? C
L(f)? ? M (f(A?D)-2)2 M (f(C?B)-3)2
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
2
3
79

• Take f that returns
• the ratings already there ( )
• similar ratings for neighbouring nodes

A ? C
L(f)? ? M (f(A?D)-2)2 M (f(C?B)-3)2 0.5
(f(A?B)-f(A?D))
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
2
3
80

• Take f that returns
• the ratings already there ( )
• similar ratings for neighbouring nodes

A ? C
L(f)? ? M (f(A?D)-2)2 M (f(C?B)-3)2 0.5
(f(A?B)-f(A?D))
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
2
3
81
A ? C
L(f)? ? M (f(A?D)-2)2 M (f(C?B)-3)2 0.5
(f(A?B)-f(A?D))
1
A ? B
0.5
1
A ? D
C ? B
M
M
2
3
82
A ? C
L(f)? ? M (f(A?D)-2)2 M (f(C?B)-3)2 0.5
(f(A-gtB)-f(A?D))
1
A ? B
Min(L(f))
0.5
1
A ? D
C ? B
M
M
2
3
83
Daniele Quercia
new graph
good rating function
?
C
1
f
2
?
A
B
84
Does LDTP work?
85
Daniele Quercia
Useful? Tested on real data (Advogato gt 55K
user ratings)
86
Daniele Quercia
Useful? Tested on real data (Advogato gt 55K
user ratings)
87
Daniele Quercia
Fast and Light?
88
Daniele Quercia
Fast and Light?
For propagating A?B (worst case) Transmit
30KB run for 2.8ms
89
LDTP
MobiRate
Use
Store
ICDM07
Ubicomp08
90
Use
Store
... are a step towards...
91
rescuing drowning users!
help!
92
producers consumers
93
producers consumers
94
future what they produce?
95
(1) Tags/Ratings (2) Movements
96
(1) Tags/Ratings (2) Movements
97
Tags/Ratings ? Content Filtering
98
(No Transcript)
99
1 month
100
1 How to suggest fresh content
Tags ? Filtering
101
?
classical
rock
102
2 How to handle niches
103
(No Transcript)
104
3 How to kill spam
105
1 fresh content
2 handle niches
3 kill spam
106
(1) Tags/Ratings (2) Movements
107
(No Transcript)
108
Who talks to whom
109
Network
110
Network
Why?
111
Changes in a company
112
1 How to accept changes
influentials
113
2 How to measure changes
114
3 How to measure social pulse
common room
115
1 accept changes
2 measure changes
3 social pulse
116
FriendSense Recommending Friends Using Mobile
Phones
117
producers consumers
(1) Tags/Ratings (2) Movements
118
ltEndgt
119
Daniele Quercia
All this on
mobblog ucl
120
(No Transcript)
121
(No Transcript)
122
Assumption ID is unique public key
123
If not unique ? Sybil attacks!
124
Sybil Attacks Against Mobile Users Friends and
Foes to the Rescue