Information Technology Reference
In-Depth Information
control parameters which should be tuned, (b) because of fixed learning rate
it lacks stability, (c) rather elaborated technique for visualizing resulting graph
must be invented.
An immune algorithm is able to generate the reference vectors (called anti-
bodies) each of which summarizes basic properties of a small group of documents
treated here as antigens
2
. This way the clusters in the immune network spanned
over the set of antibodies will serve as internal images, responsible for mapping
existing clusters in the document collection into network clusters. In essence,
this approach can be viewed as a successful instance of exemplar-based learning
giving an answer to the question ”what examples to store for use during gener-
alization, in order to avoid excessive storage and time complexity, and possibly
to improve generalization accuracy by avoiding noise and overfitting”, [17].
3.1
aiNet Algorithm for Data Clustering
The artificial immune system aiNet [7] mimics the processes of clonal selection,
maturation and apoptosis [8] observed in the natural immune system. Its aim is
to produce a set of antibodies binding a given set of antigens (i.e. documents).
The ecient antibodies form a kind of immune memory capable to bind new
antigens suciently similar to these from the training set.
Like in SOM, the antigens are repeatedly presented to the memory cells (being
matured antibodies) until a termination criterion is satisfied. More precisely, a
memory structure
M
consisting of matured antibodies is initiated randomly with
few cells. When an antigen
ag
i
is presented to the system, its anity
aff
(
ag
i
,ab
j
)
to all the memory cells is computed. The value of
aff
(
ag
i
,ab
j
) expresses how
strongly the antibody
ab
j
binds the antigen
ag
i
. From a practical point of view
aff
(
ag
i
,ab
j
) can be treated as a degree of similarity between these two cells
3
.
The greater anity
aff
(
ag
i
,ab
j
), the more stimulated
ab
j
is.
The idea of clonal selection and maturation translates into next steps (here
σ
d
,and
σ
s
are parameters). The cells which are most stimulated by the antigen
are subjected to clonal selection (i.e. each cell produces a number of copies
proportionally to the degree of its stimulation), and each clone is subjected to
mutation (the intensity of mutation is inversely proportional to the degree of
stimulation of the mother cell). Only clones
cl
which can cope successfully with
the antigen (i.e.
aff
(
ag
i
,cl
)
>σ
d
) survive. They are added to a tentative memory,
M
t
, and the process of clonal suppression starts: an antibody
ab
j
too similar to
another antibody
ab
k
(i.e.
aff
(
ab
j
,ab
k
)
>σ
s
) is removed from
M
t
. Remaining
cells are added to the global memory
M
.
These steps are repeated until all antigens are presented to the system. Next
the degree of anity between all pairs
ab
j
,ab
k
∈
M
is computed and again too
2
Intuitively by antigens we understand any substance threatening proper function-
ing of the host organism while antibodies are protein molecules produced to bind
antigens. A detailed description of these concepts can be found in [8].
3
In practical applications this measure can be derived from any metric dissimilarity
measure
dist
as
aff
(
ag
i
,ab
j
)=
d
max
−dist
(
ag
i
,ab
j
)
d
max
,where
d
max
stands for the maximal
dissimilarity between two cells.
