Information Technology Reference
In-Depth Information
antibodies that match the organism's molecules to die. Alternatively, Jerne's
idiotypic network theory
suggests that in addition to antigens, antibodies can
recognise other antibodies. The antibody-to-antibody recognition can operate
on multiple levels, forming chains of suppression and reinforcement, creating
complex reaction networks, which regulate the concentration of self-matching
antibodies [13].
AIS are particularly good at maintaining and boosting diversity. This is
achieved in two ways.
Heterostasis
, the preservation of diversity, is implicitly
accomplished using local selection mechanisms. In accordance with Menczer's
own solution for reinforcing diversity in GAs, an antibody is typically selected
to clone based on its anity to an antigen and not its relative importance (fit-
ness) with respect to the rest of the cells. Furthermore, algorithms based on
idiotypic network theory achieve diversity explicitly using suppression of simi-
lar antibodies.
Heterogenesis
, on the other hand, refers to the introduction of
diversity and is accomplished either through somatic hypermutation, or the re-
cruitment of new cells.
By combining heterostasis with heterogenesis, immune-inspired IF systems
appear well suited to the problem of profile adaptation. With heterostasis suf-
ficient coverage of the information space is achieved for the representation of
a user's multiple interests, while it is also ensured that new, previously unmet
information items (antigens) can be recognised. Heterogenesis, further facilitates
the exploration of new areas in the information space. By maintaining and boost-
ing diversity, these systems may prove effective in adapting a user profile to both
short-term variations and long-term changes in the user's interests. They may
prove advantageous, comparing to evolutionary approaches, in maintaining their
viability during adaptation.
This potential in applying AIS to the problem of profile adaptation in content-
based filtering has not been explored yet. Existing immune-inspired approaches
to IF concentrate instead on learning, in a batch mode, to discriminate between
relevant (self) and non-relevant (non-self) information items. In [14] for exam-
ple, AIS have been used for filtering computer generated graphics. Antibodies
and antigens are both modelled as 9 digit, real valued vectors and their anity
is measured as the maximum arithmetic distance between two matching digits.
[15] applied AIS to the problem of binary document classification. Antibodies
and antigens are binary keyword vectors of fixed length, where some of the bits
are masked with the special “don't care” symbol
#
. The anity between cells is
measured as the percentage of matching bits, ignoring any #. Finally, AIS have
also been applied to the task of email filtering [16]. Here antibodies and antigens
(emails) are both modelled as unweighted keyword vectors of varied length and
their anity measured as the proportion of common keywords. A similar appli-
cation is described in [17] where antibodies correspond to regular expressions
composed by randomly recombining information from a set of libraries. These
immune-inspired approaches to IF either don't deal with profile adaptation, or
treat it implicitly with periodic retraining of the profiles.
