Class-based Aggressive Feature Selection for Polynomial Networks Text Classifiers – an Empirical Study

Feature Selection (FS) is a crucial preprocessing step in Text Classification (TC) systems. FS can be either Class-Based or Corpus-Based. Polynomial Network (PN) classifiers have proved recently to be competitive in TC using a very small subset of corpora features. This paper presents an empirical study of the performance of PN classifiers using Aggressive Class-Based FS. Seven of the stateof-the art FS metrics are experimented and compared: Chi Square (CHI), Information Gain (IG), Odds Ratio (OR), GSS, NGL coefficient, Document Frequency (DF), and Gain Ratio (GR).The study is conducted on the Reuters Benchmark Corpus. Experimental results are presented in terms of both micro-averaged and macro-averaged precision, recall and F measures. Results reveal that aggressive Class-Based Chi-Square and DF metrics work best for Reuters using PN classifiers compared to the other five FS metrics experimented in this research.