ciency of the underlying algorithms. We have focused on the latter
approach by making the classifier incremental (i.e. not having to re-
train with all data every time new data arrives) and efficiently support
a large number of classes (Cyberspace services continuously generate a
large amount of user data and are plentiful with respect to the possible
navigational directions that corresponds to classes in the classifier),
making it decremental in order to efficiently "forget" outdated data
(Cyberspace services change over time), and finally by parallelizing
the algorithm in order to make it run efficiently on relatively cheaply
available PC-based parallel cluster systems (Cyberspace services are
frequently very large).
type of training is called decremental training or simply forgetting.
the-art classifiers using cross-validation (see materials and methods
chapter) and several datasets.
claims that using Java and C++ give similar computational perfor-
mance Mangione [1998]; Wilson [2003]. The main results (using paired
t-tests on results from all datasets) were that our classifier outper-
formed the WEKA classifiers by one order of magnitude or more for
the computational performance. For classification accuracy we found
no significant difference from the WEKA-classifiers, with the excep-
tion of the naive bayes classifier. The reason that our classifier is more
accurate than naive bayes is that click-stream data tends to have rela-
tions between attributes, this characteristic disputes the assumptions
of independence between attributes that naive bayes has.
on subscription and valued customer membership services. The solution is
aimed towards relaxing the restrictions of mobile devices and wireless com-
munications (Paper A).
community (cited by Chang [2002]; Turban et al. [2002]; Vrechopoulos et al.