3.2
Classification Accuracy
In order to make some more general conclusions on MIPSVM's accuracy com-
pared to the other classifiers, pairwise T-tests have been used
These tests concluded that on the 18 datasets
the default configurations of MIPSVM, Naive Bayes and C4.5 were used (ap-
pendix C). MIPSVM is significantly more accurate than Naive Bayes (5% con-
fidence level, p-value = 0.013), and not significantly different from C4.5 (even
though C4.5 is more accurate). On the 14 datasets the default configurations
of MIPSVM, Logistic Regression and Voted Perceptron were used (appendix C)
pairwise T-tests showed that MIPSVM and Logistic Regression have not signif-
icantly different accuracy (even though Logistic Regression is more accurate),
and that MIPSVM and Voted Perceptron have not significantly different accu-
racy (even though MIPSVM is more accurate).
A modestly bold conclusion and recommendation is that MIPSVM is a suit-
able alternative to Naive Bayes as the default classifier when first attacking a
classification problem.
3.3
Further Work
Opportunities for further work include:
Add support for a variable number of features in examples
Develop support for parallelized decremental PSVM
Add kernel support
Add incremental balancing mechanisms to improve accuracy for cases with
many, potentially unbalanced classes.
Investigate the performance effect of altering matrix multiplication algo-
rithms
Investigate whether the symmetric matrix system (A) can be transformed
into a Toeplitz or Hankel matrix system for more efficient computation
References
1. C. L. Blake and C. J. Merz. UCI Repository of Machine Learning Databases.
Online, http://www.ics.uci.edu/~mlearn/MLRepository.php, 1998.
2. Chih-Chung Chang and Chih-Jen Lin. LIBSVM: A Library for Support Vector Ma-
chines, 2001. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
3. Chih-Chung Chang Chih-Wei Hsu and Chih-Jen Lin. A practical guide to SVM
classification.
Department of Computer Science and Information Technology,
National Taiwan University. Paper available at http://www.csie.ntu.edu.tw/
~cjlin/papers/guide/guide.pdf.
4. H°
avard Engum. IncRidge: A Software Tool for Scalable Incremental and Decre-
mental Classifier Algorithms based on Support Vector Approximations. Technical
report, IDI, NTNU, July 2003.
Paper I
147