Empirical Study to Evaluate the Performance of Classification Algorithms on Public Datasets

Abstract

In today’s world, a huge amount of data is stored in the form of electronic documents in the World Wide Web. Text classification algorithms have been widely used for classifying those text documents into a fixed number of predefined classes. The applicable scopes and their performances of these algorithms are different. Therefore, finding an appropriate algorithm for a dataset is becoming a significant emphasis for researchers to solve practical problems quickly. This paper puts forward an experimental evaluation of five significant text classification algorithms with each other and with TF and TF-IDF feature selection methods built using decision tree (C5.0), support vector machine, K-nearest neighbor, Naïve Bayes, and neural network on four public datasets, namely 20news-bydate, ohsumed-first-20000-docs, Reuters 21578-Apte-90 Cat, and 20 Newsgroup. The experimental results are examined from multiple perspectives and summarized to provide usefulness of different algorithms on different datasets.