Measuring Behavioral Trust in Social NetworksSibel Adali, et al.
IEEE International Conference on Intelligence and
Security Informatics
Presented by: Liang ZhaoNorthern Virginia Center
OutlineIntroductionBehavior TrustTwitter dataExperiment ResultsConclusion
IntroductionTrust vs. Social NetworkEvaluate Trust in Social NetworkAssumptionsPurpose of this paper
Trust vs. Social Network
Trust → Social Network (SN)◦Forms coalitions◦Identifies influential nodes in SN◦Depicts the flow of information
Social Network → Trust◦Communities induce greater
trust◦Information flow enhances trust
Evaluate Trust in Social Network
Our own predisposition to trust.Relationship with others.Our opinions towards others.
Whether we trust others?
AssumptionsDoes not consider semantic
information.Only consider social tiesTrust is a social tie between
a trustor and trustee.Social ties can be observed
by communication behaviors.
Degree of Trust can change.
Behavior Trust: Measure of trust is based on social behavior.
Social behaviors can conversely enhanceor reduce the trust.
Purpose of this paperMeasure trust based on the
communi-cation behavior of the actors in SN.
Input:◦Communication Stream of Social
Network: {<sender, receiver, time>,…,<sender,
receiver, time>}Output:
◦Behavior trust graph Nodes: actors in SN, e.g., . Edges’ weights: strength of trust, e.g., .
Behavior TrustConversations & PropagationsConversations behavior based
◦Conversations grouping◦Conversation Trust Computation
Propagation behavior based◦Propagation Trust◦Potential Propagations Counting◦Propagation Trust Computation
Conversations & PropagationsThis paper considers two kinds of
behavior:◦ Conversations: Two nodes converse means
they are more likely to trust each other.
◦ Propagations: A propagates info from B indicates A trust B.
undirected directed
Conversations groupingThe set of messages exchanged
between A and B is: .
Average time between messages is:
Rule: two consecutive messages ,
are in the same conversation if .
𝑡1 𝑡 2 𝑡 3 𝑡 4 𝑡5 𝑡 6 𝑡7
Conversation Trust Computation
Rules:◦ Longer Conversations imply more trust.◦ More Conversations imply more trust.◦ Balanced participation between two
actors imply more trust.Trust (namely Edge’s weight in trust
graph):
Entropy function:
: the fraction sent by one actor; the fraction sent by the other actor.
Propagation Trust
Given communication statistics alone, we cannot definitely determine which messages from B are propagations from A.
So we turn to counting “potential propagations”.
𝐴
details
?
Potential Propagations CountingPotential Propagations must
satisfy the following constraint:
Matching “incoming to B” messages with “outgoing from B” messages:
𝑠1−𝑡1<𝜏𝑚𝑖𝑛𝜏𝑚𝑖𝑛<𝑠2−𝑡 1<𝜏𝑚𝑎𝑥𝑠3−𝑡 2>𝜏𝑚𝑎𝑥𝜏𝑚𝑖𝑛<𝑠3− 𝑡3<𝜏𝑚𝑎𝑥
No cross
Propagation Trust ComputationNotations:
◦ the number of propagations by B.◦ the number of potential
propagations.◦the number of messages A sent to B.
Strategy 1: Strategy 2:
The fraction of B’s energy spent on propagating A’ messages.
The fraction of A’s messages worthy to be propagated by B.
Twitter DataData Volume:
◦2M users (1.9M senders).◦230K tweets per day.
Data format:◦(sender, receiver, time).
Ground Truth Label of Trust: retweeting◦Directed
◦Broadcast
ExperimentCompute Conversation &
Propagation Graphs.Overlaps between Conversation &
Propagation Graphs.Validate Conversation &
Propagation Graphs using retweets.
Computing Conversation & Propagation Graphs
Data:◦15M Directed tweets for conversation
graph.◦34M broadcast tweets for propagation
graph.
Settings:
Computing Conversation & Propagation Graphs (continued)
To achieve comparison between conversation and propagation graphs: treat the undirected edge as two directed ones.
Overlaps between Conversation & Propagation GraphsCluster these two graphs based
on the weighted edges to discover communities:
Overlaps evaluation:
Random set of clusters with same size distribution; repeat 1000 times.
Graph validation using retweets.Assumption:
◦A retweet is a propagation.◦When a user propagates information
from some other user, there must be some element of trust between them.
◦ indicates directed trust: .◦Directed retweet is more determinative
than broadcast retweet in indicating trust.
Graph validation using retweets (contd.)Conversational Trust Graph
Validation:◦Nodes: 20% are also presented in
retweets graph.◦Edges: as follows.
: Random graph, which consists of randomly selected nodes. The edges are communications between the nodes.
: Prominence graph, which consists of most active nodes. The edges are communications between the nodes.
Graph validation using retweets (contd.)Propagation Trust Graph
Validation:◦Nodes: 20% are also presented in
retweets graph.◦Edges: as follows.
ConclusionMethod advantages:
◦ Propose a measurable behavior trust metric.
◦ Does not need semantic information.◦ Can be applied to dynamic network.◦ The proposed metric reasonably
correlate with retweets.◦ Can be applied to general social
networks other than Twitter.◦ Good scalability due to low
computational cost on statistical communication data.
Future WorksVerify the potentially casual
relationship between conversation and propagation behavior.
The intersection of conversation and propagation graphs would be a more stringent measure of trust.
Improve the purity of trust measurement by considering semantics of messages.
Trust should be dependent on context (e.g., we trust a doctor in medical science, but not necessarily in finance analysis.
Improve the trust measurement by considering the quality and value of messages.
Thank you