1 networking 2012 parallel and distributed systems group, delft university of technology, the...

16
1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized Interaction-Based Reputation Systems Dimitra Gkorou, Tamás Vinkó, Nitin Chiluka, Johan Pouwelse, and Dick Epema

Upload: percival-pearson

Post on 17-Jan-2018

213 views

Category:

Documents


0 download

DESCRIPTION

3 NETWORKING 2012 Reputation Systems: Basic Concepts the goal of reputation in large scale systems: establish trust among users incentives for good behavior Interaction-Based Reputation Systems why not use the complete history? resource requirements: computation + storage capacity dynamic behavior: population turnover + changing behavior reputation algorithm complete history of interactions reputations of nodes

TRANSCRIPT

Page 1: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

1NETWORKING 2012

Parallel and Distributed Systems Group,Delft University of Technology, the Netherlands

May 22, 2012

Reducing the History in Decentralized

Interaction-Based Reputation SystemsDimitra Gkorou, Tamás Vinkó, Nitin Chiluka, Johan Pouwelse, and Dick Epema

Page 2: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

2NETWORKING 2012

Overview

• Interaction-based Reputation Systems• Limitations of the Complete History • Reducing the History• Evaluation• Conclusion

Page 3: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

3NETWORKING 2012

Reputation Systems: Basic Concepts

• the goal of reputation in large scale systems:• establish trust among users• incentives for good behavior

• Interaction-Based Reputation Systems

• why not use the complete history?• resource requirements: computation + storage capacity• dynamic behavior: population turnover + changing

behavior

reputation algorithm

complete history of

interactionsreputations

of nodes

Page 4: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

4NETWORKING 2012

• Complete History (CH): modeled as a growing directed weighted graph

• Reduced History (RH): a dynamically maintained subset of CH with fixed size

• removal of the least important nodes and edges in

Reducing the History: Basic Approach

Complete History (CH)

Reduced History (RH)

node removal: freshnessactivity levelreputation position

edge removal: freshnessweightposition

Page 5: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

5NETWORKING 2012

• parameters indicating the importance:• freshness (node/edge): capturing the dynamics of the

system• position (node/edge): keeping the graph connected• activity level (node): maintaining informative nodes• reputation (node): maintaining trustworthy information• weight (edge): importance of an edge

• combined to a priority score for each node and edge

Reducing the History: Priority Score

Complete History (CH)

Reduced History (RH)

Page 6: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

6NETWORKING 2012

Reduced History: Construction

Complete History: • add new node + its edges• add new edge connecting existing nodesReduced History (fixed size): • add the new node + remove the node with the

lowest priority score• add the new edge + remove the edge with the

lowest priority scorewad

wab

wbc wd

Complete History

wad

wab

wbcwd

Reduced History

wedwed

wce wg

b

ad

b

a

dc c

ee

f fg

gwfg wfg

wge wgewfe wfe

Page 7: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

7NETWORKING 2012

Experiment Setup: Synthetic Graphs• CH growing up to 5000 nodes• random graphs:

• new nodes/edges connected to existent nodes with a constant probability

• scale-free graphs:• new nodes/edges connected to existent nodes with a

probability proportional to their degree• multiple edges correspond to weights

random graph scale-free graph

Page 8: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

8NETWORKING 2012

Experiment Setup: Real-world Graphs• Bartercast Reputation mechanism:

• Tribler: the BitTorrent P2P file-sharing system• provides incentives for contribution• peers locally store the history of their own interactions +

interactions among other peers• information exchange: using an epidemic protocol

• Bartercast graph:• crawled the Bartercast reputation mechanism (4 months)• union of all local graphs• vertices: the peers of Tribler• weighted edges: the amount of the transferred data

between two peers

Page 9: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

9NETWORKING 2012

Experiment Setup: real-world graphs• Author-to- author citation graph:

• derived from papers published in Physical Review E• vertices: the authors of papers• weighted edges: number of citations between authors

• small-world graphs• Citation graph more densely connected than Bartercast

graph # nodes

# edges aver. path length

c.c.

Bartercast

10,634 31,624 2.64 0.00074

Citation 15,360 365,319 3.29 0.11

Page 10: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

10NETWORKING 2012

Computation of Reputation

• Max-flow based computation:• reputation computation of Bartercast• the weights of edges graph as flows• starting from the most central node j• reputation of peer a: the difference of flows faj and fja

• Eigenvector centrality:• well-studied metric• interactions with highly reputed nodes contribute

more• Pagerank

e

j

c

b

Wca\ac

wbi

f

wgk

kwbj

g

wjg

wfg

wge

Wba\abwga

awia

i

Page 11: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

11NETWORKING 2012

Evaluation Metrics

• the ranking of reputations is more important than their actual values

• the identification of the highest ranked nodes is more important

• consider the sequences of ranked nodes in CH and RH according to their reputation

• two metrics• ranking error: the minimum number of swaps needed to get

the same ranking sequence in RH and CH• ranking overlap: the fraction of common nodes in the

sequences of top-raked nodes in RH and CH

Page 12: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

12NETWORKING 2012

Evaluation: ranking errorMax-Flow Pagerank

rank

ing

erro

r

Size of RH relatively to the size of CH Growth of CH relatively to the size of RH

Max-Flow Pagerank

• scale-free and real-world graphs exhibit smaller ranking error

• pagerank exhibits smaller ranking error

Page 13: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

13NETWORKING 2012

Evaluation: ranking overlap

Max-FlowRa

nkin

g ov

erla

pRa

nkin

g ov

erla

p

Size of RH relatively to the size of CH

Pagerank

• max-flow achieves much higher ranking overlap

• random graphs exhibit the worst ranking overlap

Page 14: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

14NETWORKING 2012

Evaluation: ranking overlap

Max-FlowRa

nkin

g ov

erla

pRa

nkin

g ov

erla

p

Pagerank

Growth of CH relatively to the size of RH

• max-flow achieves much higher again ranking overlap

Page 15: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

15NETWORKING 2012

Conclusions

• the performance of RH depends on the topology• scale-free and real-world graphs exhibit smaller ranking

error and higher ranking overlap

• the performance of RH depends on the reputation algorithm• pagerank achieves lower ranking error• max-flow achieves higher ranking overlap

• RH achieves good accuracy for real-world graphs

Page 16: 1 NETWORKING 2012 Parallel and Distributed Systems Group, Delft University of Technology, the Netherlands May 22, 2012 Reducing the History in Decentralized

16NETWORKING 2012

Questions?

www.pds.ewi.tudelft.nl

www.tribler.org

contact: [email protected]