QCon London: Mastering long-running processes in modern architectures
Sas web 2010 lora-aroyo
1. Trust and Reputation in Social
Internetworking Systems
Lora Aroyo1
Pasquale De Meo1
Domenico Ursino2
1VU University Amsterdam, the Netherlands
2DIMET – University of Reggio Calabria, Italy
2. Social Networks Added Value
! advertise products and disseminate innovations
& knowledge
! find information relevant to users
! find relevant users, e.g. LinkedIn
! spread opinions, e.g., personal, social or
political
! interesting for:
! museums, broadcaster, government institutions
3. Online Identities
! Increasing number of identities
! different information sharing tasks
! connect with different communities
! UK adults have ~1.6 online profiles
! 39% of those with one profile have at
least two other profiles
! Companies exploring the potential of
social internetworking
! Platform(s) for data portability among
social networks
5. What’s Needed?
! mechanisms to:
! help users find reliable users
! disclose malicious users or spammers
! stimulate the level of user participation
! deal with trust in linked data
! deal with different contexts and policies for
accessing, publishing and re-distributing data
6. What’s the Goal?
! model to represent Social Internetworking
components & their relationships
! understand Social Internetworking structural
properties and see how it differs from traditional
social networks
! model to compute
trust & reputation based on
linked data
7. Requirements
! trust should be tied to user’s performance, i.e., providing
beneficial contributions to other users
! consider that users are involved in a range of activities, e.g.,
tagging, posting comments, rating
! represent a wide range of heterogeneous entities, e.g. users,
resources, posts, comments, ratings and their interactions
(vs. single role nodes in graphs)
! edges need to support n-ary relationships vs. binary in graphs
! multi-dimensional network vs. one-dimensional in graphs
! easy to manipulate and intuitive model
8. Graph-Based Approaches
! Model user community as graph G
! edges reflect explicit trust relationship between
users
! G is sparse, thus often need for inferring trust
values
! model trust & reputation in force-mass-
acceleration style capture all factors and
combine them in a set of equations
! resulting model is too complicated to be handled
9. Link-Based Approaches
! link analysis algorithms, e.g. PageRank or
HITS, model trust as a measure of system
performance, e.g., number of corrupted files in a
peer of a P2P network
! attack-resistant to manipulate reputation score
! model trust & reputation in force-mass-
acceleration style capture all factors and
combine them in a set of equations
! resulting model is too complicated to be handled
10. SIS Approach
! Social Graph API (list of public URLs and
connections for person p (e.g., Twitter page of p and
contacts of p)
! Hypergraph
! nodes labels with
object role
! multiple hyperedges
between two nodes
! hyperedges – link two
or more entities
11. SIS Pilot: Analysis
! We gathered from multiple social networks, e.g.,
LiveJournal, Twitter, Flickr:
! 1, 252, 908 user accounts
! 30, 837, 012 connections between users
! The probability P(k) that a user has created an
account in k networks is distributed as:
P(k) ~ k-4.003
! Few users are affiliated to multiple networks
! More than 90% of users are affiliated to less than
3 networks
12. Canonization Procedures
! Map gathered data to graph with following
properties:
! High network modularity, i.e., nodes tend to form
dense clusters with few inter-cluster edges
! Small world phenomenon, i.e., paths between
arbitrary pairs of nodes are usually short
13. Reputation in SIS
! Setting:
! users post resources &rate resources posted by others
! To compute reputation we assume that:
! User-high-reputation if he authors high quality resources
! Resource-high-quality if it gets a high average rating &
posted by users with high reputation
! mutual reinforcement principle
14. Trust in SIS
! n = # of users in SIS m = # of resources they authored
! r(i) = reputation of useri q(j) = quality of resourcej
! e(j) = average rating of resourcej
! Aij = 1 if useri posted a resourcej and Aij = 0 otherwise
! r = Aq and q = AT r + e r = (I – AAT)-1Ae
.
! compute dominant eigenvector of a symmetric matrix
! easy to compute even if A gets large (AT = transpose of A
and I = nxn identity matrix)
15. Future Work
! Gather a larger amount of data to analyze further the
structural properties of SIS
! Test the effectiveness of the approach for reputation
computing
! Test with real users in the social space of Agora (Social
Event-based History browsing) and in PrestoPrime
(Social Semantic Taging)
.
! Ontology-based model of trust and reputation in
different domains (with LOD)