1 Whats Up: P2P news recommender Anne-Marie Kermarrec Joint work with Antoine Boutet, Davide Frey (INRIA) and Rachid Guerraoui (EPFL) Gossple workshop.

1

What’s Up:P2P news recommender

Anne-Marie Kermarrec

Joint work with Antoine Boutet,Davide Frey (INRIA) and Rachid Guerraoui (EPFL)

Gossple workshop 2010

Gossple workshop 2010 2

The social Web

Web content is generated by you, me, your friends and millions of others

The Web has turned social

Content comes from everywhere

Gossple workshop 2010 3

Is it equally relevant?

Gossple workshop 2010 4

Is it equally relevant?

Gossple workshop 2010 5

Is it equally relevant?

Gossple workshop 2010 6

What’s wrong with news feed?

Amazon recommends me a fryer

Some of my Facebook write in Italian

LeMonde.fr wants to inform me on the Champion’s ligue

Gossple workshop 2010 7

Why is it so difficult?

• Even a space restricted to users explicit subscriptions is too large a database

• Dynamic• Recommendations not always user-centric• Explicit links not always that relevant• Classical pub/sub do not filter enough

Granularity of a user seems too coarse

Gossple workshop 2010 8

Cascading over explicit links

Gossple workshop 2010 9

Fine grain tuning calls fordecentralisation

Gossple workshop 2010 10

What’s up

• Decentralised information dissemination channel

• Simple interface: I like it or I don’t

• Exploit implicit social links

Gossple workshop 2010 11

An implicit pub/sub system

Gossple workshop 2010 12

What’s up in a nutshell

Gossple workshop 2010 13

What’s up challenges

Gossple workshop 2010 14

•Who are my social acquaintances

•How to discover them?

•How to disseminate news ?

Similarity metric

Through gossip

Biased epidemic protocol

What’s up: Gossple net

Gossple workshop 2010 15

What’s up challenges

Gossple workshop 2010 16

•Who are my social acquaintances

•How to discover them?

•How to disseminate news ?

Similarity metric

An implicit social network

Gossple workshop 2010 17

Which nodes should be considered as social acquaintances?

Model• U(sers) × I(tems) (news)• Profile(u) = vector of liked news• Minimal information

Similarity metrics• Overlap

• Cosine similarity

• Multi-interest similarity

Gossple workshop 2010 18

What’s up challenges

Gossple workshop 2010 21

•Who are my social acquaintances

•How to discover them?

•How to disseminate news ?

Through gossip

22

The Gossple network

Gossple workshop 2010

Copyright: E. Rivière

Gossip similarity protocol.

Gossip-based peer sampling service

Gossple workshop 2010 24

Building the social network

• Two gossip protocols• Similarity-based Peer Sampling• Random Peer Sampling

• When p encounters q• Evaluate potential new view, based on set

similarity metric• Use of Bloom filters to limit the communication

overhead

RPS

SPS

RPS

SPS

What’s up in a nutshell

Gossple workshop 2010 25

What’s up challenges

Gossple workshop 2010 26

•Who are my social acquaintances

•How to discover them?

•How to disseminate news ? Biased epidemic protocol

Dissemination

Gossple workshop 2010 27

Heterogeneous

Homogeneous

HeterogeneousHomogeneous

Involvement (fanout)

Expectations

EpidemicDissemination

F=log(N)

HeterogeneousGossip

F≈ log(N) on average

28

BEEP: orientation and amplification

Orientation: to whom?

Gossple workshop 2010

Forwardto

friends

Forwardto

random

Amplification: to how many?

Increase fanout

Decreasefanout

Beep: I like it

Gossple workshop 2010 29

I like it!

Beep: I don’t

Gossple workshop 2010 30

I dislike it!

Tuning BEEP

• Orientation• The news carries the list of visited users• A profile: sum of interests of users who liked it

• Amplification✔ F≈ log(N) friends✔ Amplification depends on the similarity between the

news and the user✖ F≈ 1 or 2 random

Gossple workshop 2010 31

Evaluation

• User Metrics• Spam• Recall• Precision

• System metric• Number of messages• Redundancy (useless messages)

• Traces• Synthetic clustered traces• Real dataset: 700 Digg users/2000 news/1 week

Gossple workshop 2010 32

Preliminary results

Gossple workshop 2010 33

Algorithm Precision Recall Spam

Perfect 1 1 0

Gossip fanout=log(n)=7

0.28 0.94 0.74

Cascading through explicit friends from Digg

0.39 0.71 0.71

WhatsUp fanout=11/1 ; ttl=12

0. 52 0.6

WhatsUp without no social users

To take away

• Automatic light news recommender

• Analysis through mean field theory

• Experimental evaluation

Next: diversity of sources, trust, privacy

Gossple workshop 2010 34

Thank you

35

www.gossple.fr

Gossple workshop 2010

Gossple workshop 2010 36