Irrelevant Feature and Rule Removal for Structural Associative Classification Using Structure-Preserving Flat Representation - Feature Selection for Data and Pattern Recognition - page 205

Sect. 10.3.3 . In this section, an illustrative example is provided using an Academic
Institution WebLogs example as reference. The DSM is reflected in the structure
of T 0 in Table 10.4 and the corresponding tree is shown in Fig. 10.3 (Session 0).
Transaction T 0 becomes the general structure of DSM and the header in Table 10.7
to reflect the attributes names. Every transaction that remains in the T db will be
matched against the DSM and every node label placed in the matching column (i.e.
under the matching node ( x i ) in the DSM). The flat data format of T db fromTable 10.4
is illustrated in Table 10.7 .
10.3.5 Representing Disconnected Trees w.r.t. DSM
As discussed earlier in Sect. 10.3.3 , the rules from DSM can be converted into pre-
order string encoding of the subtrees, and hence are represented as subtrees of the
tree database. However, some rules may not be representatives of valid subtrees.
For example, it is possible that some items in the rules correspond to sibling nodes
in the original tree, while the parent or any ancestor node connecting those in the
original tree is not present in the rules discovered using DSM approach. Hence, this
would result in an invalid subtree as the nodes are disconnected. In addressing this
matter, one can add the other nodes that make it into a valid subtree but flag them
as irrelevant. The process consists of sequentially listing the values of each matched
node in DSM, while retaining the level of embedding information of each current
node in DSM and in the subtree pattern. Since the DSM itself is ordered according
to the pre-order traversal, this results in pre-order string encodings of the subtrees.
As a simple illustrative example, consider the following associations/patterns
extracted from an Academic Institution WebLogs Data:
P 1 : business-intelligence human-space-computing phd-msc,
P 2 : scholarships management phd-a-msc.
With respect to pattern ( P 1 )inFig. 10.4 and pattern ( P 2 )inFig. 10.5 , the items
(nodes) in the rule correspond to sibling nodes in the original tree, while the parent
or any ancestor node connecting those in the original tree is not present in the rule.
Hence, this would result in an invalid subtree as the nodes are disconnected. This is
illustrated in both Figs. 10.4 and 10.5 , where irrelevant nodes are shaded grey. One
can also choose to display the labels of nodes that are there to contextualize the infor-
mation, i.e. scholarships and management and phd-a-msc, which would essentially
contextualize the specific rule constraints. Additionally, the labels of nodes can be
displayed in order to contextualize the information in the tree. In this work, these
rules are recognized as FullTree rules.