Upload
daniele-quercia
View
3.567
Download
0
Tags:
Embed Size (px)
DESCRIPTION
How to offer digital content to mobile users by combining tagging with reputation systems
Citation preview
daniele quercia & licia capra & valentina zanardi
I’m doing my PhD @
University
College
London
<MobiSys>
We research distributed sys
mobblog ucl
We blog
Web 2.0(mobile: $22.4bn)
Location?
The next big thing!
$8.1 bn
$8.1 bn
This talk is about tools for...
Consuming content on the move
Content Creation: By publishers?
People
Mainly: people
People
Mainly: people
People
create a ...
Digital Tapestry
Creation:Distributed!
Consumption:Centralized!
Why?
Money!
Making Money? No, short of ideas!
Wisdom of the (Programming)Crowd
API
wired freedom by APIs
Creation:Distributed!
What if...
Consume: centralized decentralized
Existing Tools for Internet
Tools for decentralized consumption
<1> Query matching<2> Ratings for sources
<1> Query matching
Centralized Solutions
similarity(query,item)
queryitem
similarity(query,item)
queryitem
Coverage (digg out content) +Accuracy (no good content)-
Idea behind SocialRanking
[RecSys08] Social Ranking:Finding Relevant Content in Web 2.0
sim(query,item) +
sim(query,item) + sim(issuer,tagger)
issuer tagger
similarity(issuer,tagger)
issuer tagger
similarity(issuer,tagger)
Coverage (digg out content)
Accuracy (good content)++
Future:Decentralized!
<2> Ratings for sources
Store & Use & Categories
Store & Use & Categories
How to store ratings?
1.Log(credentials)
2. Gossip(to check each credential)
1.Log(credentials)
2. Gossip(to check each credential)
Impractical
Idea behind MobiRate
[Ubicomp08] MobiRate: Making Mobile Raters Stick to their Word
1.Sealed Log(of credentials)
2. Gossip(to check seals only)
1.Sealed Log(of credentials)
2. Gossip(to check seals only)
Practical
works?
Security:It outperforms existing solutions
“heaviest” protocol runs < 2sec
“longest” protocol completedin 2.5ms (if Bluetooth 100kb/s)
Store & Use & Categories
Use ratings to make predictions
Daniele Quercia
Traditional way: Trust propagation ?A B
C
Daniele Quercia
That way works onWeb & “binary” ratings
Idea behind LDTP
[ICDM07] Lightweight Distributed Trust Propagation
Daniele Quercia
1
?A B
C2
Daniele Quercia
1
?A B
C2
1 2
?
new graph
AB
AC CBf
Daniele Quercia
1
?A B
C2
1 2
?
new graph
AB
AC CB
“good” rating function
f
works?
Daniele Quercia
Useful? Tested on real data (Advogato: > 55K user ratings)
Daniele Quercia
Useful? Tested on real data (Advogato: > 55K user ratings)
Daniele Quercia
Fast and “Light”?
Daniele Quercia
Fast and “Light”?
For propagating AB (worst case)
Transmit 30KB& run for 2.8ms
Store & Use & Categories
Greek Coins
Roman Coins
Coins Chairs
Antiques
universal ontology
Idea behind TRULLO
[MobiQuitous07] TRULLO - local trust bootstrapping for ubiquitous devices
Daniele Quercia
Users learn from their ratings
Daniele Quercia
Users learn from their ratings
How?
Daniele Quercia
Singular
Value
Decomposition
Daniele Quercia
SVD
works?
Daniele Quercia
Good “porting” upon few ratings
Daniele Quercia
Nokia 3230
Daniele Quercia
Daniele Quercia
Store & Use & Categories
Issues: many!
Issue 1: Evaluation
Issue 2: Privacy
If you are looking for Kinky boots
whether you are …
a woman…
or a man a woman…
Thorny problem:
… or a woman…How to keep it secret
Issue 2: Privacy
mobblog ucl
All this on ...