Information Technology Reference
In-Depth Information
that in time dependent optimisation problems, like profile adaptation, where the
optimum, or optima, change over time, GAs suffer due to their elitist character.
GAs converge and there is a progressive loss in diversity as the optimum prolifer-
ates and spreads over the population. This can cause evolutionary IF systems to
specialise to one area (topic) of interest and reach a state which inhibits further
adaptation.
The immune system's ability to discriminate between the host organism's own
molecules (self) and foreign, possibly harmful, molecules (non-self), serves well
as a metaphor to the problem of IF. Typically, Immune-inspired IF systems
employ a dynamic repertoire of profile representations (antibodies) that learn to
discriminate between relevant information (self) and non-relevant information
(non-self). The composition of this repertoire changes in a way that, in contrast
to GAs, not only maintains, but also boosts diversity. As we further discuss in
the next section, this characteristic may prove advantageous when dealing with
adaptive IF. Despite this potential however, the application of immune-inspired
approaches to the problem of adapting the user profile to changes in the user's
multiple interests has not been fully explored yet. Existing immune-inspired IF
systems concentrate on traditional routing applications where profiles are trained
in a batch mode and then used for filtering. Profile adaptation is either ignored,
or treated simply by periodically repeating the training process.
To explore the application of immune-inspired ideas to the problem of profile
adaptation we evaluate in this paper Nootropia
1
, a user profiling model that has
been introduced in [4,5]. The immune network is used as a metaphor to build a
network of terms that represents a user's multiple interests (section 3.1) and that
adapts to changes in them through a process of self-organisation (section 3.2).
The evaluation methodology (section 4) uses virtual users to simulate a variety
of interest changes. The results show that through self-organisation a user profile
that represents more than one topic of interest can adapt to both modest and
radical interest changes. They exhibit the profile's ability both to ”learn” and
to ”forget” and signify the importance of the network structure during this
process. The evaluation methodology itself is of interest because it reflects more
accurately than existing standards the multimodal and time-dependent nature
of adaptive IF. The current work is part of ongoing research on biologically
inspired IF that seeks to compare AIS and GAs on this challenging problem.
2
Evolutionary and Immune-Inspired IF
The insight behind GAs is that the fundamental components of biological evo-
lution can be used to evolve solutions to problems within computers. They are
stochastic search techniques that have been traditionally applied to optimisation
problems. Typically in evolutionary IF a population of profiles, which collec-
tively represent the user interests, is maintained [6,7,8,9]. The population evolves
according to user feedback. Individual profiles that better represent the user
1
Greek word for: “an individual's or a group's particular way of thinking, someone's
characteristics of intellect and perception”.
