(e.g. pre-fetching based on click-stream prediction from usage logs), and due
to the services large scale and continuously changing nature classifiers need to
handle this, this motivates the parallel, incremental and decremental classifiers
proposed.
(C4.5, Logistic Regression, Voted Perceptron) on general classification data sets,
user click-streams from an actual web usage log and a synthetic game usage log
from the developed MMOG simulator. The incremental classifier showed to be
about 1 order of magnitude (or more) faster than the classifiers compared with,
and significantly more accurate than the naive bayes classifier on the selected data
sets. We didn't find any significant difference between the classification accuracy
of the proposed classifier and the other classifiers on the selected data sets; this
may suggest that the proposed classifiers can be useful for user prediction in
cyberspace services.
for a live service. The peer-to-peer collaborative filtering architecture can
potentially be implemented and tested for the recommendation of mobile
phone ringing tones and backgrounds.
and potentially test whether user-positioning data can be used to improve
the click-stream predictions.
proved intelligence support (e.g. BDI and automatic planning), simulation
of intra-game e-commerce activities (e.g. trade between players). The in-
creased realism can then be used to test various approaches for logging and
data mining of player behavior.
[1994]), and GP (Koza [1998]).