Information Technology Reference
In-Depth Information
user marks viewed documents as either relevant, or not. Changes in the user
interests are reflected by changes in the content of documents that received
user feedback. Nootropia's adaptation in response to user feedback is achieved
through a process that is described in detail in [5]. In accordance with Varela's
view, the process is not antigen driven. Relevant documents and their terms
correspond to the production of antibodies by the bone marrow, to which the
immune network reacts structurally.
In summary, the process comprises five deterministic steps. Given a document
that received positive feedback, terms in the document are weighed and those
with weight over a certain threshold are extracted. Some of the extracted terms
already appear in the profile (immune repertoire) and some are new. In the sec-
ond step, we employ a local selection scheme. Each of the profile terms that is also
contained in the document, is selected and its weight (concentration) is increased
by its weight in the document. Terms are not selected based on their relative im-
portance. The overall additional weight is subsequently subtracted evenly from
all profile terms. Therefore, during this step, the overall profile weight remains
constant, but a redistribution of weight towards terms that appear in the rel-
evant document takes place. These variations in the weight (concentration) of
terms (antibodies) correspond to the networks dynamics. They cause changes in
the ordering of terms. Eventually, some terms loose their weight. These terms
are removed from the profile and the sum of their initial weight, i.e. the weight
of each new profile term, is evenly subtracted from the remaining terms. The
fourth step of the process involves the new extracted terms. These are added to
the profile with initial weight equal to their weight in the relevant document.
The recruitment and removal of terms (antibodies) implements the network's
metadynamics. Finally, links between existing and new terms are generated and
the weight of existing links is updated. It is important to note that due to the
way document evaluation is performed the survival of a newly recruited term
depends not only on its initial weight, but also I¿I
2
the current network struc-
ture and I¿I
2
the term's place in it. In other words endogenous selection takes
place since it is the network itself which selects those terms that will survice.
In the case of a document that received negative feedback, only the first three
of the above five steps take place. The process differs in that during the second
step, the weight of profile terms that have been extracted from the document
gets decremented, rather than incremented, by their weight in the document.
The overall subtracted weight is then equally divided among all profile terms.
According to [5] the above process allows the profile to adapt through self-
organisation of the network's structure. Hierarchies that correspond to topics
that received positive feedback grow, while those that did not receive positive
feedback decline. Such variations in the size of hierarchies can allow the profile
to quickly adapt to short-term variations in user interests. In a similar way,
more substantial long term changes can also be accounted for. A new hierar-
chy may develop when a new topic of interest emerges. On the other hand, a
hierarchy that corresponds to a topic that is no longer interesting progressively
disintegrates and is eventually forgotten. Negative feedback is not essential for
