background image
Empirical Comparison of Accuracy and
Performance for the MIPSVM classifier with
Existing Classifiers
Amund Tveit
Division of Intelligent Systems
Department of Computer and Information Science,
Norwegian University of Science and Technology,
N-7491 Trondheim, Norway
amundt@idi.ntnu.no
Abstract. This paper presents an empirical comparison of the Multicat-
egory Incremental Proximal Support Vector Machine Classifier (MIPSVM)
against the C4.5, Naive Bayes, Voted Perceptron, SMO, SVM and Logis-
tic Regression classifiers on several datasets. The datasets are from the
UCI Machine Learning repository, a Web Usage Log, and game usage
logs from the Zereal Massively Multiplayer Online Game Simulator.
MIPSVM is found to be close to 1 order of magnitude (or more) faster
than the other classifiers in all experiments. Based on pairwise T-test
comparisons of accuracy between MIPSVM and the other classifiers,
MIPSVM was found to be significantly more accurate than the Naive
Bayes classifier, and not having significantly different accuracy than the
other classifiers on the tested data sets.
1
Introduction
The objective of this paper is to test both computational performance and clas-
sification accuracy of Multicategory Incremental Proximal Support Vector Ma-
chine classifier (MIPSVM) [8] against other classifiers on several datasets.
All experiments are performed using the average accuracy from 10-fold cross-
validation method, i.e. training on 9/10s of the data and test on the remaining
1/10 for all 1/10s in the dataset (i.e. 10 tests per dataset).For classifation ac-
curacy comparison, graphs of average accuracy percentage from 10-fold cross
validation are used, similar for time comparison
Computational performance is measured in average cross-validation wallclock
time shown on a logarithmic scale relative to MIPSVM. MIPSVM has a value
of 1 (hence not visible!), so if another classifier algorithm has a value of 10 it
means it is one order of magnitude (ten times) slower than MIPSVM.
Finally we analyze the results using pairwise T-tests to get a general indica-
tion of MIPSVMs accuracy relative to the other classifiers compared with.
136
Empirical Comparison of MIPSVM with existing classifiers

<< - < - > - >>