Comparison of Classifiers According to Number of Trials


Korkmaz O. E., Aydemir Ö.

2nd International Symposium on Applied Sciences and Engineering (ISASE2021), Erzurum, Türkiye, 7 - 09 Nisan 2021, ss.285-289

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Erzurum
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.285-289
  • Atatürk Üniversitesi Adresli: Evet

Özet

Pattern recognition and machine learning based studies have attracted considerable interest in recent years. Because of the fact that there is no a fixed feature and classification method that can solve all kinds of pattern recognition and machine learning problem, in studies determining the most appropriate feature and classifier together with its parameters requires research and is a timeconsuming process. Comparison of classifiers from different perspectives gives the researchers an idea about the solution of their problems. In the literature, while classifiers were compared in different feature distributions, sizes or classification performance criteria, there were limited studies about classifiers which compared in terms of number of trials which was an important parameter especially for researchers. In this study, Support Vector Machines, k Nearest Neighbor, and Linear Discrimination Analysis classifiers for four different feature spaces were compared in terms of different trial counts. The results revealed that the classification accuracy for different distribution types and classifiers changed depending on the number of trials. Furthermore, as the number of trials increased, the standard deviation decreased. It means that the stability increased as expected with the increase in the number of trials. It can be said that the k Nearest Neighbor algorithm generally provides superior classification accuracy for all types of distributions.