Information Technology Reference
In-Depth Information
Term Weight
Term Weight
DT1
DT2
term
activated term
symmetric link
DT3
direction of
energy spreading
Fig. 1.
Hierarchical Term Network: (left) deactivated, (right) activated
the document, and
w
i
the weight of an activated term
t
i
. Alternatively, additional
evidence derived from the pattern of activation can also be taken into account. In
equation 2,
S
1(
D
) is complemented with the additional factor
log
(1 + (
b
+
d
)
/b
).
We call
b
the “breadth” of the document, that can be estimated as the number
of activated terms that did not disseminate any energy (fig. 1(right): terms DT1,
DT2 and DT3).
d
stands for the “depth” of the document and is estimated as the
number of activated terms that disseminated energy. Hence,
S
2(
D
) which has
been adopted for our experiments, awards documents which activate connected
subnetworks and not isolated terms.
S
1(
D
)=
i∈A
w
i
·
E
i
log
(
NT
)
(1)
log
(1 +
b
+
d
b
S
2(
D
)=
S
1
D
·
)
(2)
This directed spreading activation process takes into account the term depen-
dencies that the network represents to establish non-linear document evaluation.
How much a term contributes to a document's relevance score depends not only
on its weight, but also on the term's place within the hierarchy and its links to
other terms. It depends on the current network structure. This is a property of
the model that distinguishes it from traditional approaches to IF, like the vector
space model, which ignore term dependencies. It has been argued and supported
experimentally that it is this property which allows the effective representation
of multiple topics with a single user profile [4].
3.2
Profile Adaptation
Once the user profile is initialised its life cycle begins and can be used to evaluate
documents. Based on the assigned relevance scores and an appropriate thresh-
old a distinction can be made between relevant (self) and non-relevant (non-self)
documents. Nevertheless, documents are typically presented to the user in de-
creasing relevance order and is left to the user to decide which documents to
read. The user expresses satisfaction, or dissatisfaction, of the filtering results
through relevance feedback. Here we only consider binary feedback
3
where the
3
Scaled feedback is also possible.
