Feature Selection Strategy in Text Classification

Traditionally, the best number of features is determined by the so-called "rule of thumb", or by using a separate validation dataset. We can neither find any explanation why these lead to the best number nor do we have any formal feature selection model to obtain this number. In this paper, we conduct an in-depth empirical analysis and argue that simply selecting the features with the highest scores may not be the best strategy. A highest scores approach will turn many documents into zero length, so that they cannot contribute to the training process. Accordingly, we formulate the feature selection process as a dual objective optimization problem, and identify the best number of features for each document automatically. Extensive experiments are conducted to verify our claims. The encouraging results indicate our proposed framework is effective.