Databases Reference
In-Depth Information
rule mining approach and the randomised selector based rule mining approach,
whereas a significant difference in times spent on mining significant rules can
be seen.
6 Conclusion
This chapter is concerned with an investigation of CARM. A overview of
existing CARM algorithms was provided in Sect. 2 where five existing rule
weighting schemes used in CARM algorithms were reviewed. A rule weighting
scheme was proposed in Sect. 3 that was used to distinguish the significant
CARs from the insignificant ones. Consequently a rule ordering strategy was
proposed, based on the “best first” case satisfaction approach, which can be
applied when classifying “unseen” data. The concept of selectors [19] was
summarised in Sect. 3 together with some discussion of the randomised se-
lectors. A novel rule mining approach was presented in Sect. 4 based on the
concepts of selectors (both deterministic and randomised). In theory, the pro-
posed rule mining approach identifies significant CARs in time
O
(
k
2
n
2
)in
its deterministic fashion, and
O
(
kn
) in its randomised fashion. This mining
approach avoids computing CAR scores and finding significant CARs on a
“one-by-one” basis, which will require an exponential time
O
(2
n
). In Sect. 5,
two sets of evaluations were presented that evidence:
1. The proposed rule weighting and rule ordering approach's perform well
with respect to the accuracy of classification.
2. The proposed randomised rule mining approach is comparable to the
“one-by-one” rule mining approach in significant CAR identification with
respect to both the accuracy of classification and the e
ciency of compu-
tation.
From the experimental results, it can be seen that the accuracy of clas-
sification obtained by the proposed randomised rule mining approach can be
better than the accuracy obtained by the “one-by-one” approach. Further re-
search is suggested to identify improved rule weighting scheme to find more
significant rules in
R
. Other obvious direction for further research include:
finding other rule ordering mechanisms that give a better classification ac-
curacy; investigating other techniques to replace the proposed deterministic
and/or randomised selectors to give a better performance; etc.
Acknowledgement
The authors would like to thank Professor Leszek Gasieniec and Professor Paul
Leng of the Department of Computer Science at the University of Liverpool
for their support with respect to the work described here.
