Immune-Inspired Adaptive Information Filtering - Artificial Immune Systems - page 422

3Noo rop a

Nootropia is the first attempt to apply ideas drawn from the immune system
to the problem of profile adaptation to changes in a user's multiple interests.
The model is described in detail in [5], where it is argued that Nootropia has
common characteristics with computational models of the immune system that
have been developed within the the context of Maturana and Varela's autopoietic
theory [18]. According to Varela the immune system is not antigen driven, but
instead, an organisationally closed network that reacts autonomously in order to
define and preserve the organism's identity, in what is called self-assertion [19].
This is achieved through two types of change: variation in the concentration of
antibodies called the dynamics of the system, and, on the other hand, a slower
recruitment of new cells (produced by the bone marrow) and removal of existing
cells, called the metadynamics . While dynamics play the role of reinforcement
learning, the metadynamics function as a distributed control mechanism which
aims at maintaining the viability of the network through an on-going shift in
the immune repertoire [20]. One significant aspect of the immune network's
metadynamics is that the system itself is responsible for selecting new cells for
recruitment in what is called endogenous selection . In this paper, we only briefly
describe Nootropia's profile representation and adaptation and concentrate in-
stead on the model's evaluation.

3.1 Profile Representation

Inspired by Varela's view of the immune network, in Nootropia, a term network
is used to represent a user's multiple interests. This profile representation is de-
scribed in detail in [4], along with a process for initialising the network based on
a set of documents that are relevant to the user. It is depicted in figure 1(left).
Terms in the network correspond to antibodies and links denote antibody-to-
antibody recognition. A term's weight corresponds to the antibody's concentra-
tion and measures how important the term is regarding the user's interests. A
link's weight on the other hand, corresponds to the a nity between antibod-
ies and measures the statistical dependencies that exist between semantically
and syntactically correlated terms. Terms in the network are ordered according
to decreasing weight. This gives rise to separate term hierarchies, one for each
general topic that is discussed in the relevant documents (e.g. two overlapping
topics in fig. 1(left)). This is a significant transformation that is the basis for the
non-linear evaluation of documents according to the represented topics.
More specifically, when confronted with a new document D ,profileterms
that appear in D are activated (fig 1(right)). Subsequently, each activated term
disseminates part of its activation to activated terms higher in the hierarchy
that it is linked to. The amount of activation that is disseminated between two
activated terms is proportional to the weight of the corresponding link.
It is then possible to calculate the document's relevance score based on the
final activation of activated terms. In the simplest case, this is done using equa-
tion 1, where A is the set of activated profile terms, NT the number of terms in