Some modifications should be done to achieve better performance than the previous one:
===================================================================================================
About the machine learning methods:
(1) The libsvm deal with the multiple class problem with 1-against-1, however, even though they mention that the 1-against-1 is compariable to the results of 1-against-the rest, we still want to have a try
http://www.csie.ntu.edu.tw/~cjlin/libsvm/faq.html#f419
The famous paper discuss the one-against-one and one-against-the-rest. http://www.csie.ntu.edu.tw/~cjlin/papers/multisvm.pdf
some conclusions: http://hal.inria.fr/docs/00/10/39/55/PDF/cr102875872670.pdf
For problems with few
classes, like digit recognition, the “one against all”
strategy seems significantly more accurate; while for
problems with more classes, like Latin letter recognition,
the difference of accuracy between the two strategies
seems much less significant. Lastly, for problems with a
very large number of classes, like Chinese or Japanese
ideogram recognition, we suspect that the unbalance of
the number of the samples causes problem with the “one
against all” strategy, especially when it has few training
samples per class. Moreover, the “one against one”
strategy, which is more modular, is more suitable for
speeding up the decision making process by combining
with other classifiers [15]. Furthermore, if the number of
training samples is very large, the training time can
libsvm 的FAQ: http://blog.sina.com.cn/s/blog_490236f10100042s.html
(2) Except using SVM to do classification, maybe decision tree methods could also be ustilized.