Trust and Technology - Trust Theory: A Socio-Cognitive and Computational Model - page 346

the relationship between physical magnitudes of stimuli and the perceived intensity of the
stimuli). This transformation also allows the idea that the average opinion yields the lower
certainty of transformed trust. It helps to reduce the subjectivity in opinion-based datasets so
that the evidence-based approaches like Cert-Prop can apply.
The attention to the mathematical properties of the operators, fails sometimes to catch
the deeper nature of the trust phenomenon. For example, when I receive different, diverging
evaluations (about Y ) from two or more agents (say J and Z ), not only can I discount the
degree of trust in Y on the basis of J or Z 's trustworthiness and believability in me (Hang
et al. , 2009), not only do I have to combine those converging or diverging values with some
mathematical 'aggregation', but I have to choose among different heuristics , strategies. Not
necessarily is the final value of trust a mix of the various values. I may, for example, be a very
suspicious and prudent guy (or adopt a prudent strategy), and, although J and Z say that Y is
sure and good, since W says that Y is not good (or sure) I adopt W 's view, and put aside J
and Z 's evaluations. Or I might have an optimistic attitude and adopt always the best, more
favorable estimation. Or I might have a strong esteem of Z (as evaluator) and trust him very
much; although J and W have different opinions I do not care about them, I adopt Z 's opinion
(trust) without discounting it by combining it with the other evaluations. In sum, there are
different possible heuristics in combining (or not) various evaluations; there is not a unique
('rational') equation.
In (Richardson et al. , 2003) the trust propagation model allows each user to maintain trust
in a small number of other users. This method first enumerates all paths between the user and
every other user who has a local belief in a given statement. Then, the belief associated with
each path (concatenation operator) is calculated, and combined with the beliefs associated
with all paths (aggregation operator). The aggregation operator is the same as the Cert-Prop's
one while the concatenation operator is different.
Trust metrics 1 compute quantitative estimates of how much trust an agent X should have in
Y , taking into account trust ratings from other agents on the network.
Two main important applications of trust metrics are: Advogate (Levien, 2009) and Apple-
seed (Ziegler, 2009). Both these metrics can be classified as local group trust metrics . Local
is intended versus Global : where Global take into account all peers in the network and the
links connecting them; while Local trust metrics take into account personal bias. They operate
on partial trust graph information (the web of trust for an agent X is the set of relationships
emanating from X and passing through nodes X (directly or indirectly) trusts.
Advogate computes a set of accepted nodes in three steps. First, it is assigned a capacity to
every node as a function of the shortest path distance from the seed to that node. Second, there
is a transformation of the graph, adding extra edges from each node to a special node (called
supersink). Third, it is computs the maximum network flow for the new graph: the accepted
nodes are those that have a flow across the special node (supersink).
In contrast to Advogate , Appleseed uses spreading activation (Quilian, 1968). It spreads
energy across the graph, and when propagated through a node, divides energy among sucessors
based on the edge weights. The main idea in Appleseed is to simulate the spectral decomposition
and it requires several iterations to converge towards the set of acceptable nodes.
1 Here considered with a different meaning with respect to the previous section 'Different kinds of metrics in this
chapter.