Feature selection via maximizing global information gain for text classification

Feature selection is a vital preprocessing step for text classification task used to solve the curse of dimensionality problem. Most existing metrics (such as information gain) only evaluate features individually but completely ignore the redundancy between them. This can decrease the overall discriminative power because one feature's predictive power is weakened by others. On the other hand, though all higher order algorithms (such as mRMR) take redundancy into account, the high computational complexity renders them improper in the text domain. This paper proposes a novel metric called global information gain (GIG) which can avoid redundancy naturally. An efficient feature selection method called maximizing global information gain (MGIG) is also given. We compare MGIG with four other algorithms on six datasets, the experimental results show that MGIG has better results than others methods in most cases. Moreover, MGIG runs significantly faster than the traditional higher order algorithms, which makes it a proper choice for feature selection in text domain.